Upload folder using huggingface_hub
Browse files- README.md +171 -0
- Whisper_Transcription_+_NeMo_Diarization.ipynb +942 -0
- audio_clean.py +126 -0
- audio_cleaning_test.py +57 -0
- autodiarize.py +255 -0
- bulktranscript.py +222 -0
- combinesets.py +94 -0
- concat.py +21 -0
- consolidate_datasets.py +110 -0
- diarize.py +227 -0
- diarize_parallel.py +203 -0
- helpers.py +392 -0
- mergefolders.py +56 -0
- nemo_msdd_configs/diar_infer_general.yaml +92 -0
- nemo_msdd_configs/diar_infer_meeting.yaml +92 -0
- nemo_msdd_configs/diar_infer_telephonic.yaml +92 -0
- nemo_process.py +29 -0
- process_srt_wav.py +80 -0
- test.py +269 -0
- test2.py +201 -0
- test3.py +214 -0
- transcription_helpers.py +73 -0
- update_metadata.py +30 -0
- youtube_to_wav.py +33 -0
README.md
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Autodiarize
|
2 |
+
|
3 |
+
This repository provides a comprehensive set of tools for audio diarization, transcription, and dataset management. It leverages state-of-the-art models like Whisper, NeMo, and wav2vec2 to achieve accurate results.
|
4 |
+
|
5 |
+
## Table of Contents
|
6 |
+
|
7 |
+
- [Installation](#installation)
|
8 |
+
- [Usage](#usage)
|
9 |
+
- [Diarization and Transcription](#diarization-and-transcription)
|
10 |
+
- [Bulk Transcription](#bulk-transcription)
|
11 |
+
- [Audio Cleaning](#audio-cleaning)
|
12 |
+
- [Dataset Management](#dataset-management)
|
13 |
+
- [YouTube to WAV Conversion](#youtube-to-wav-conversion)
|
14 |
+
- [LJSpeech Dataset Structure](#ljspeech-dataset-structure)
|
15 |
+
- [Contributing](#contributing)
|
16 |
+
- [License](#license)
|
17 |
+
|
18 |
+
## Installation
|
19 |
+
|
20 |
+
### 1. Clone the repository:
|
21 |
+
|
22 |
+
```bash
|
23 |
+
git clone https://github.com/your-username/whisper-diarization.git
|
24 |
+
cd whisper-diarization
|
25 |
+
```
|
26 |
+
|
27 |
+
### 2. Create a Python virtual environment and activate it:
|
28 |
+
|
29 |
+
```bash
|
30 |
+
./create-env.sh
|
31 |
+
source autodiarize/bin/activate
|
32 |
+
```
|
33 |
+
or if you want to ruin your python env
|
34 |
+
|
35 |
+
### Install the required packages:
|
36 |
+
|
37 |
+
```bash
|
38 |
+
pip install -r requirements.txt
|
39 |
+
```
|
40 |
+
|
41 |
+
## Usage
|
42 |
+
|
43 |
+
### Diarization and Transcription
|
44 |
+
|
45 |
+
The `diarize.py` script performs audio diarization and transcription on a single audio file. It uses the Whisper model for transcription and the NeMo MSDD model for diarization.
|
46 |
+
|
47 |
+
```bash
|
48 |
+
python diarize.py -a <audio_file> [--no-stem] [--suppress_numerals] [--whisper-model <model_name>] [--batch-size <batch_size>] [--language <language>] [--device <device>]
|
49 |
+
```
|
50 |
+
|
51 |
+
- `-a`, `--audio`: Path to the target audio file (required).
|
52 |
+
- `--no-stem`: Disables source separation. This helps with long files that don't contain a lot of music.
|
53 |
+
- `--suppress_numerals`: Suppresses numerical digits. This helps the diarization accuracy but converts all digits into written text.
|
54 |
+
- `--whisper-model`: Name of the Whisper model to use (default: "medium.en").
|
55 |
+
- `--batch-size`: Batch size for batched inference. Reduce if you run out of memory. Set to 0 for non-batched inference (default: 8).
|
56 |
+
- `--language`: Language spoken in the audio. Specify None to perform language detection (default: None).
|
57 |
+
- `--device`: Device to use for inference. Use "cuda" if you have a GPU, otherwise "cpu" (default: "cuda" if available, else "cpu").
|
58 |
+
|
59 |
+
### Bulk Transcription
|
60 |
+
|
61 |
+
The `bulktranscript.py` script performs diarization and transcription on multiple audio files in a directory.
|
62 |
+
|
63 |
+
```bash
|
64 |
+
python bulktranscript.py -d <directory> [--no-stem] [--suppress_numerals] [--whisper-model <model_name>] [--batch-size <batch_size>] [--language <language>] [--device <device>]
|
65 |
+
```
|
66 |
+
|
67 |
+
- `-d`, `--directory`: Path to the directory containing the target files (required).
|
68 |
+
- Other arguments are the same as in `diarize.py`.
|
69 |
+
|
70 |
+
### Audio Cleaning
|
71 |
+
|
72 |
+
The `audio_clean.py` script cleans an audio file by removing silence and applying EQ and compression.
|
73 |
+
|
74 |
+
```bash
|
75 |
+
python audio_clean.py <audio_path> <output_path>
|
76 |
+
```
|
77 |
+
|
78 |
+
- `<audio_path>`: Path to the input audio file.
|
79 |
+
- `<output_path>`: Path to save the cleaned audio file.
|
80 |
+
|
81 |
+
### Dataset Management
|
82 |
+
|
83 |
+
The repository includes several scripts for managing datasets in the LJSpeech format.
|
84 |
+
|
85 |
+
#### Merging Folders
|
86 |
+
|
87 |
+
The `mergefolders.py` script allows you to merge two LJSpeech-like datasets into one.
|
88 |
+
|
89 |
+
```bash
|
90 |
+
python mergefolders.py
|
91 |
+
```
|
92 |
+
|
93 |
+
Follow the interactive prompts to select the directories to merge and specify the output directory.
|
94 |
+
|
95 |
+
#### Consolidating Datasets
|
96 |
+
|
97 |
+
The `consolidate_datasets.py` script consolidates multiple LJSpeech-like datasets into a single dataset.
|
98 |
+
|
99 |
+
```bash
|
100 |
+
python consolidate_datasets.py
|
101 |
+
```
|
102 |
+
|
103 |
+
Modify the `base_folder` and `output_base_folder` variables in the script to specify the input and output directories.
|
104 |
+
|
105 |
+
#### Combining Sets
|
106 |
+
|
107 |
+
The `combinesets.py` script combines multiple enumerated folders within an LJSpeech-like dataset into a chosen folder.
|
108 |
+
|
109 |
+
```bash
|
110 |
+
python combinesets.py
|
111 |
+
```
|
112 |
+
|
113 |
+
Enter the name of the chosen folder when prompted. The script will merge the enumerated folders into the chosen folder.
|
114 |
+
|
115 |
+
### YouTube to WAV Conversion
|
116 |
+
|
117 |
+
The `youtube_to_wav.py` script downloads a YouTube video and converts it to a WAV file.
|
118 |
+
|
119 |
+
```bash
|
120 |
+
python youtube_to_wav.py [<youtube_url>]
|
121 |
+
```
|
122 |
+
|
123 |
+
- `<youtube_url>`: (Optional) URL of the YouTube video to download and convert. If not provided, the script will prompt for the URL.
|
124 |
+
|
125 |
+
## LJSpeech Dataset Structure
|
126 |
+
|
127 |
+
The `autodiarize.py` script generates an LJSpeech-like dataset structure for each input audio file. Here's an example of how the dataset structure looks:
|
128 |
+
|
129 |
+
```
|
130 |
+
autodiarization/
|
131 |
+
├── 0/
|
132 |
+
│ ├── speaker_0/
|
133 |
+
│ │ ├── speaker_0_001.wav
|
134 |
+
│ │ ├── speaker_0_002.wav
|
135 |
+
│ │ ├── ...
|
136 |
+
│ │ └── metadata.csv
|
137 |
+
│ ├── speaker_1/
|
138 |
+
│ │ ├── speaker_1_001.wav
|
139 |
+
│ │ ├── speaker_1_002.wav
|
140 |
+
│ │ ├── ...
|
141 |
+
│ │ └── metadata.csv
|
142 |
+
│ └── ...
|
143 |
+
├── 1/
|
144 |
+
│ ├── speaker_0/
|
145 |
+
│ │ ├── speaker_0_001.wav
|
146 |
+
│ │ ├── speaker_0_002.wav
|
147 |
+
│ │ ├── ...
|
148 |
+
│ │ └── metadata.csv
|
149 |
+
│ ├── speaker_1/
|
150 |
+
│ │ ├── speaker_1_001.wav
|
151 |
+
│ │ ├── speaker_1_002.wav
|
152 |
+
│ │ ├── ...
|
153 |
+
│ │ └── metadata.csv
|
154 |
+
│ └── ...
|
155 |
+
└── ...
|
156 |
+
```
|
157 |
+
|
158 |
+
Each input audio file is processed and assigned an enumerated directory (e.g., `0/`, `1/`, etc.). Within each enumerated directory, there are subdirectories for each speaker (e.g., `speaker_0/`, `speaker_1/`, etc.).
|
159 |
+
|
160 |
+
Inside each speaker's directory, the audio segments corresponding to that speaker are saved as individual WAV files (e.g., `speaker_0_001.wav`, `speaker_0_002.wav`, etc.). Additionally, a `metadata.csv` file is generated for each speaker, containing the metadata for each audio segment.
|
161 |
+
|
162 |
+
The `metadata.csv` file has the following format:
|
163 |
+
|
164 |
+
```
|
165 |
+
filename|speaker|text
|
166 |
+
speaker_0_001|Speaker 0|Transcribed text for speaker_0_001
|
167 |
+
speaker_0_002|Speaker 0|Transcribed text for speaker_0_002
|
168 |
+
...
|
169 |
+
```
|
170 |
+
|
171 |
+
Each line in the `metadata.csv` file represents an audio segment, with the filename (without extension), speaker label, and transcribed text separated by a pipe character (`|`).
|
Whisper_Transcription_+_NeMo_Diarization.ipynb
ADDED
@@ -0,0 +1,942 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"attachments": {},
|
5 |
+
"cell_type": "markdown",
|
6 |
+
"metadata": {
|
7 |
+
"colab_type": "text",
|
8 |
+
"id": "view-in-github"
|
9 |
+
},
|
10 |
+
"source": [
|
11 |
+
"<a href=\"https://colab.research.google.com/github/MahmoudAshraf97/whisper-diarization/blob/main/Whisper_Transcription_%2B_NeMo_Diarization.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
|
12 |
+
]
|
13 |
+
},
|
14 |
+
{
|
15 |
+
"attachments": {},
|
16 |
+
"cell_type": "markdown",
|
17 |
+
"metadata": {
|
18 |
+
"id": "eCmjcOc9yEtQ"
|
19 |
+
},
|
20 |
+
"source": [
|
21 |
+
"# Installing Dependencies"
|
22 |
+
]
|
23 |
+
},
|
24 |
+
{
|
25 |
+
"cell_type": "code",
|
26 |
+
"execution_count": null,
|
27 |
+
"metadata": {
|
28 |
+
"id": "Tn1c-CoDv2kw"
|
29 |
+
},
|
30 |
+
"outputs": [],
|
31 |
+
"source": [
|
32 |
+
"!pip install git+https://github.com/m-bain/whisperX.git@a5dca2cc65b1a37f32a347e574b2c56af3a7434a\n",
|
33 |
+
"!pip install --no-build-isolation nemo_toolkit[asr]==1.21.0\n",
|
34 |
+
"!pip install git+https://github.com/facebookresearch/demucs#egg=demucs\n",
|
35 |
+
"!pip install deepmultilingualpunctuation\n",
|
36 |
+
"!pip install wget pydub\n",
|
37 |
+
"!pip install --force-reinstall torch torchaudio torchvision\n",
|
38 |
+
"!pip uninstall -y nvidia-cudnn-cu12\n",
|
39 |
+
"!pip install numba==0.58.0"
|
40 |
+
]
|
41 |
+
},
|
42 |
+
{
|
43 |
+
"cell_type": "code",
|
44 |
+
"execution_count": null,
|
45 |
+
"metadata": {
|
46 |
+
"id": "YzhncHP0ytbQ"
|
47 |
+
},
|
48 |
+
"outputs": [],
|
49 |
+
"source": [
|
50 |
+
"import os\n",
|
51 |
+
"import wget\n",
|
52 |
+
"from omegaconf import OmegaConf\n",
|
53 |
+
"import json\n",
|
54 |
+
"import shutil\n",
|
55 |
+
"from faster_whisper import WhisperModel\n",
|
56 |
+
"import whisperx\n",
|
57 |
+
"import torch\n",
|
58 |
+
"from pydub import AudioSegment\n",
|
59 |
+
"from nemo.collections.asr.models.msdd_models import NeuralDiarizer\n",
|
60 |
+
"from deepmultilingualpunctuation import PunctuationModel\n",
|
61 |
+
"import re\n",
|
62 |
+
"import logging\n",
|
63 |
+
"import nltk\n",
|
64 |
+
"from whisperx.alignment import DEFAULT_ALIGN_MODELS_HF, DEFAULT_ALIGN_MODELS_TORCH\n",
|
65 |
+
"from whisperx.utils import LANGUAGES, TO_LANGUAGE_CODE"
|
66 |
+
]
|
67 |
+
},
|
68 |
+
{
|
69 |
+
"attachments": {},
|
70 |
+
"cell_type": "markdown",
|
71 |
+
"metadata": {
|
72 |
+
"id": "jbsUt3SwyhjD"
|
73 |
+
},
|
74 |
+
"source": [
|
75 |
+
"# Helper Functions"
|
76 |
+
]
|
77 |
+
},
|
78 |
+
{
|
79 |
+
"cell_type": "code",
|
80 |
+
"execution_count": null,
|
81 |
+
"metadata": {
|
82 |
+
"id": "Se6Hc7CZygxu"
|
83 |
+
},
|
84 |
+
"outputs": [],
|
85 |
+
"source": [
|
86 |
+
"punct_model_langs = [\n",
|
87 |
+
" \"en\",\n",
|
88 |
+
" \"fr\",\n",
|
89 |
+
" \"de\",\n",
|
90 |
+
" \"es\",\n",
|
91 |
+
" \"it\",\n",
|
92 |
+
" \"nl\",\n",
|
93 |
+
" \"pt\",\n",
|
94 |
+
" \"bg\",\n",
|
95 |
+
" \"pl\",\n",
|
96 |
+
" \"cs\",\n",
|
97 |
+
" \"sk\",\n",
|
98 |
+
" \"sl\",\n",
|
99 |
+
"]\n",
|
100 |
+
"wav2vec2_langs = list(DEFAULT_ALIGN_MODELS_TORCH.keys()) + list(\n",
|
101 |
+
" DEFAULT_ALIGN_MODELS_HF.keys()\n",
|
102 |
+
")\n",
|
103 |
+
"\n",
|
104 |
+
"whisper_langs = sorted(LANGUAGES.keys()) + sorted(\n",
|
105 |
+
" [k.title() for k in TO_LANGUAGE_CODE.keys()]\n",
|
106 |
+
")\n",
|
107 |
+
"\n",
|
108 |
+
"\n",
|
109 |
+
"def create_config(output_dir):\n",
|
110 |
+
" DOMAIN_TYPE = \"telephonic\" # Can be meeting, telephonic, or general based on domain type of the audio file\n",
|
111 |
+
" CONFIG_FILE_NAME = f\"diar_infer_{DOMAIN_TYPE}.yaml\"\n",
|
112 |
+
" CONFIG_URL = f\"https://raw.githubusercontent.com/NVIDIA/NeMo/main/examples/speaker_tasks/diarization/conf/inference/{CONFIG_FILE_NAME}\"\n",
|
113 |
+
" MODEL_CONFIG = os.path.join(output_dir, CONFIG_FILE_NAME)\n",
|
114 |
+
" if not os.path.exists(MODEL_CONFIG):\n",
|
115 |
+
" MODEL_CONFIG = wget.download(CONFIG_URL, output_dir)\n",
|
116 |
+
"\n",
|
117 |
+
" config = OmegaConf.load(MODEL_CONFIG)\n",
|
118 |
+
"\n",
|
119 |
+
" data_dir = os.path.join(output_dir, \"data\")\n",
|
120 |
+
" os.makedirs(data_dir, exist_ok=True)\n",
|
121 |
+
"\n",
|
122 |
+
" meta = {\n",
|
123 |
+
" \"audio_filepath\": os.path.join(output_dir, \"mono_file.wav\"),\n",
|
124 |
+
" \"offset\": 0,\n",
|
125 |
+
" \"duration\": None,\n",
|
126 |
+
" \"label\": \"infer\",\n",
|
127 |
+
" \"text\": \"-\",\n",
|
128 |
+
" \"rttm_filepath\": None,\n",
|
129 |
+
" \"uem_filepath\": None,\n",
|
130 |
+
" }\n",
|
131 |
+
" with open(os.path.join(data_dir, \"input_manifest.json\"), \"w\") as fp:\n",
|
132 |
+
" json.dump(meta, fp)\n",
|
133 |
+
" fp.write(\"\\n\")\n",
|
134 |
+
"\n",
|
135 |
+
" pretrained_vad = \"vad_multilingual_marblenet\"\n",
|
136 |
+
" pretrained_speaker_model = \"titanet_large\"\n",
|
137 |
+
" config.num_workers = 0 # Workaround for multiprocessing hanging with ipython issue\n",
|
138 |
+
" config.diarizer.manifest_filepath = os.path.join(data_dir, \"input_manifest.json\")\n",
|
139 |
+
" config.diarizer.out_dir = (\n",
|
140 |
+
" output_dir # Directory to store intermediate files and prediction outputs\n",
|
141 |
+
" )\n",
|
142 |
+
"\n",
|
143 |
+
" config.diarizer.speaker_embeddings.model_path = pretrained_speaker_model\n",
|
144 |
+
" config.diarizer.oracle_vad = (\n",
|
145 |
+
" False # compute VAD provided with model_path to vad config\n",
|
146 |
+
" )\n",
|
147 |
+
" config.diarizer.clustering.parameters.oracle_num_speakers = False\n",
|
148 |
+
"\n",
|
149 |
+
" # Here, we use our in-house pretrained NeMo VAD model\n",
|
150 |
+
" config.diarizer.vad.model_path = pretrained_vad\n",
|
151 |
+
" config.diarizer.vad.parameters.onset = 0.8\n",
|
152 |
+
" config.diarizer.vad.parameters.offset = 0.6\n",
|
153 |
+
" config.diarizer.vad.parameters.pad_offset = -0.05\n",
|
154 |
+
" config.diarizer.msdd_model.model_path = (\n",
|
155 |
+
" \"diar_msdd_telephonic\" # Telephonic speaker diarization model\n",
|
156 |
+
" )\n",
|
157 |
+
"\n",
|
158 |
+
" return config\n",
|
159 |
+
"\n",
|
160 |
+
"\n",
|
161 |
+
"def get_word_ts_anchor(s, e, option=\"start\"):\n",
|
162 |
+
" if option == \"end\":\n",
|
163 |
+
" return e\n",
|
164 |
+
" elif option == \"mid\":\n",
|
165 |
+
" return (s + e) / 2\n",
|
166 |
+
" return s\n",
|
167 |
+
"\n",
|
168 |
+
"\n",
|
169 |
+
"def get_words_speaker_mapping(wrd_ts, spk_ts, word_anchor_option=\"start\"):\n",
|
170 |
+
" s, e, sp = spk_ts[0]\n",
|
171 |
+
" wrd_pos, turn_idx = 0, 0\n",
|
172 |
+
" wrd_spk_mapping = []\n",
|
173 |
+
" for wrd_dict in wrd_ts:\n",
|
174 |
+
" ws, we, wrd = (\n",
|
175 |
+
" int(wrd_dict[\"start\"] * 1000),\n",
|
176 |
+
" int(wrd_dict[\"end\"] * 1000),\n",
|
177 |
+
" wrd_dict[\"word\"],\n",
|
178 |
+
" )\n",
|
179 |
+
" wrd_pos = get_word_ts_anchor(ws, we, word_anchor_option)\n",
|
180 |
+
" while wrd_pos > float(e):\n",
|
181 |
+
" turn_idx += 1\n",
|
182 |
+
" turn_idx = min(turn_idx, len(spk_ts) - 1)\n",
|
183 |
+
" s, e, sp = spk_ts[turn_idx]\n",
|
184 |
+
" if turn_idx == len(spk_ts) - 1:\n",
|
185 |
+
" e = get_word_ts_anchor(ws, we, option=\"end\")\n",
|
186 |
+
" wrd_spk_mapping.append(\n",
|
187 |
+
" {\"word\": wrd, \"start_time\": ws, \"end_time\": we, \"speaker\": sp}\n",
|
188 |
+
" )\n",
|
189 |
+
" return wrd_spk_mapping\n",
|
190 |
+
"\n",
|
191 |
+
"\n",
|
192 |
+
"sentence_ending_punctuations = \".?!\"\n",
|
193 |
+
"\n",
|
194 |
+
"\n",
|
195 |
+
"def get_first_word_idx_of_sentence(word_idx, word_list, speaker_list, max_words):\n",
|
196 |
+
" is_word_sentence_end = (\n",
|
197 |
+
" lambda x: x >= 0 and word_list[x][-1] in sentence_ending_punctuations\n",
|
198 |
+
" )\n",
|
199 |
+
" left_idx = word_idx\n",
|
200 |
+
" while (\n",
|
201 |
+
" left_idx > 0\n",
|
202 |
+
" and word_idx - left_idx < max_words\n",
|
203 |
+
" and speaker_list[left_idx - 1] == speaker_list[left_idx]\n",
|
204 |
+
" and not is_word_sentence_end(left_idx - 1)\n",
|
205 |
+
" ):\n",
|
206 |
+
" left_idx -= 1\n",
|
207 |
+
"\n",
|
208 |
+
" return left_idx if left_idx == 0 or is_word_sentence_end(left_idx - 1) else -1\n",
|
209 |
+
"\n",
|
210 |
+
"\n",
|
211 |
+
"def get_last_word_idx_of_sentence(word_idx, word_list, max_words):\n",
|
212 |
+
" is_word_sentence_end = (\n",
|
213 |
+
" lambda x: x >= 0 and word_list[x][-1] in sentence_ending_punctuations\n",
|
214 |
+
" )\n",
|
215 |
+
" right_idx = word_idx\n",
|
216 |
+
" while (\n",
|
217 |
+
" right_idx < len(word_list)\n",
|
218 |
+
" and right_idx - word_idx < max_words\n",
|
219 |
+
" and not is_word_sentence_end(right_idx)\n",
|
220 |
+
" ):\n",
|
221 |
+
" right_idx += 1\n",
|
222 |
+
"\n",
|
223 |
+
" return (\n",
|
224 |
+
" right_idx\n",
|
225 |
+
" if right_idx == len(word_list) - 1 or is_word_sentence_end(right_idx)\n",
|
226 |
+
" else -1\n",
|
227 |
+
" )\n",
|
228 |
+
"\n",
|
229 |
+
"\n",
|
230 |
+
"def get_realigned_ws_mapping_with_punctuation(\n",
|
231 |
+
" word_speaker_mapping, max_words_in_sentence=50\n",
|
232 |
+
"):\n",
|
233 |
+
" is_word_sentence_end = (\n",
|
234 |
+
" lambda x: x >= 0\n",
|
235 |
+
" and word_speaker_mapping[x][\"word\"][-1] in sentence_ending_punctuations\n",
|
236 |
+
" )\n",
|
237 |
+
" wsp_len = len(word_speaker_mapping)\n",
|
238 |
+
"\n",
|
239 |
+
" words_list, speaker_list = [], []\n",
|
240 |
+
" for k, line_dict in enumerate(word_speaker_mapping):\n",
|
241 |
+
" word, speaker = line_dict[\"word\"], line_dict[\"speaker\"]\n",
|
242 |
+
" words_list.append(word)\n",
|
243 |
+
" speaker_list.append(speaker)\n",
|
244 |
+
"\n",
|
245 |
+
" k = 0\n",
|
246 |
+
" while k < len(word_speaker_mapping):\n",
|
247 |
+
" line_dict = word_speaker_mapping[k]\n",
|
248 |
+
" if (\n",
|
249 |
+
" k < wsp_len - 1\n",
|
250 |
+
" and speaker_list[k] != speaker_list[k + 1]\n",
|
251 |
+
" and not is_word_sentence_end(k)\n",
|
252 |
+
" ):\n",
|
253 |
+
" left_idx = get_first_word_idx_of_sentence(\n",
|
254 |
+
" k, words_list, speaker_list, max_words_in_sentence\n",
|
255 |
+
" )\n",
|
256 |
+
" right_idx = (\n",
|
257 |
+
" get_last_word_idx_of_sentence(\n",
|
258 |
+
" k, words_list, max_words_in_sentence - k + left_idx - 1\n",
|
259 |
+
" )\n",
|
260 |
+
" if left_idx > -1\n",
|
261 |
+
" else -1\n",
|
262 |
+
" )\n",
|
263 |
+
" if min(left_idx, right_idx) == -1:\n",
|
264 |
+
" k += 1\n",
|
265 |
+
" continue\n",
|
266 |
+
"\n",
|
267 |
+
" spk_labels = speaker_list[left_idx : right_idx + 1]\n",
|
268 |
+
" mod_speaker = max(set(spk_labels), key=spk_labels.count)\n",
|
269 |
+
" if spk_labels.count(mod_speaker) < len(spk_labels) // 2:\n",
|
270 |
+
" k += 1\n",
|
271 |
+
" continue\n",
|
272 |
+
"\n",
|
273 |
+
" speaker_list[left_idx : right_idx + 1] = [mod_speaker] * (\n",
|
274 |
+
" right_idx - left_idx + 1\n",
|
275 |
+
" )\n",
|
276 |
+
" k = right_idx\n",
|
277 |
+
"\n",
|
278 |
+
" k += 1\n",
|
279 |
+
"\n",
|
280 |
+
" k, realigned_list = 0, []\n",
|
281 |
+
" while k < len(word_speaker_mapping):\n",
|
282 |
+
" line_dict = word_speaker_mapping[k].copy()\n",
|
283 |
+
" line_dict[\"speaker\"] = speaker_list[k]\n",
|
284 |
+
" realigned_list.append(line_dict)\n",
|
285 |
+
" k += 1\n",
|
286 |
+
"\n",
|
287 |
+
" return realigned_list\n",
|
288 |
+
"\n",
|
289 |
+
"\n",
|
290 |
+
"def get_sentences_speaker_mapping(word_speaker_mapping, spk_ts):\n",
|
291 |
+
" sentence_checker = nltk.tokenize.PunktSentenceTokenizer().text_contains_sentbreak\n",
|
292 |
+
" s, e, spk = spk_ts[0]\n",
|
293 |
+
" prev_spk = spk\n",
|
294 |
+
"\n",
|
295 |
+
" snts = []\n",
|
296 |
+
" snt = {\"speaker\": f\"Speaker {spk}\", \"start_time\": s, \"end_time\": e, \"text\": \"\"}\n",
|
297 |
+
"\n",
|
298 |
+
" for wrd_dict in word_speaker_mapping:\n",
|
299 |
+
" wrd, spk = wrd_dict[\"word\"], wrd_dict[\"speaker\"]\n",
|
300 |
+
" s, e = wrd_dict[\"start_time\"], wrd_dict[\"end_time\"]\n",
|
301 |
+
" if spk != prev_spk or sentence_checker(snt[\"text\"] + \" \" + wrd):\n",
|
302 |
+
" snts.append(snt)\n",
|
303 |
+
" snt = {\n",
|
304 |
+
" \"speaker\": f\"Speaker {spk}\",\n",
|
305 |
+
" \"start_time\": s,\n",
|
306 |
+
" \"end_time\": e,\n",
|
307 |
+
" \"text\": \"\",\n",
|
308 |
+
" }\n",
|
309 |
+
" else:\n",
|
310 |
+
" snt[\"end_time\"] = e\n",
|
311 |
+
" snt[\"text\"] += wrd + \" \"\n",
|
312 |
+
" prev_spk = spk\n",
|
313 |
+
"\n",
|
314 |
+
" snts.append(snt)\n",
|
315 |
+
" return snts\n",
|
316 |
+
"\n",
|
317 |
+
"\n",
|
318 |
+
"def get_speaker_aware_transcript(sentences_speaker_mapping, f):\n",
|
319 |
+
" previous_speaker = sentences_speaker_mapping[0][\"speaker\"]\n",
|
320 |
+
" f.write(f\"{previous_speaker}: \")\n",
|
321 |
+
"\n",
|
322 |
+
" for sentence_dict in sentences_speaker_mapping:\n",
|
323 |
+
" speaker = sentence_dict[\"speaker\"]\n",
|
324 |
+
" sentence = sentence_dict[\"text\"]\n",
|
325 |
+
"\n",
|
326 |
+
" # If this speaker doesn't match the previous one, start a new paragraph\n",
|
327 |
+
" if speaker != previous_speaker:\n",
|
328 |
+
" f.write(f\"\\n\\n{speaker}: \")\n",
|
329 |
+
" previous_speaker = speaker\n",
|
330 |
+
"\n",
|
331 |
+
" # No matter what, write the current sentence\n",
|
332 |
+
" f.write(sentence + \" \")\n",
|
333 |
+
"\n",
|
334 |
+
"\n",
|
335 |
+
"def format_timestamp(\n",
|
336 |
+
" milliseconds: float, always_include_hours: bool = False, decimal_marker: str = \".\"\n",
|
337 |
+
"):\n",
|
338 |
+
" assert milliseconds >= 0, \"non-negative timestamp expected\"\n",
|
339 |
+
"\n",
|
340 |
+
" hours = milliseconds // 3_600_000\n",
|
341 |
+
" milliseconds -= hours * 3_600_000\n",
|
342 |
+
"\n",
|
343 |
+
" minutes = milliseconds // 60_000\n",
|
344 |
+
" milliseconds -= minutes * 60_000\n",
|
345 |
+
"\n",
|
346 |
+
" seconds = milliseconds // 1_000\n",
|
347 |
+
" milliseconds -= seconds * 1_000\n",
|
348 |
+
"\n",
|
349 |
+
" hours_marker = f\"{hours:02d}:\" if always_include_hours or hours > 0 else \"\"\n",
|
350 |
+
" return (\n",
|
351 |
+
" f\"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}\"\n",
|
352 |
+
" )\n",
|
353 |
+
"\n",
|
354 |
+
"\n",
|
355 |
+
"def write_srt(transcript, file):\n",
|
356 |
+
" \"\"\"\n",
|
357 |
+
" Write a transcript to a file in SRT format.\n",
|
358 |
+
"\n",
|
359 |
+
" \"\"\"\n",
|
360 |
+
" for i, segment in enumerate(transcript, start=1):\n",
|
361 |
+
" # write srt lines\n",
|
362 |
+
" print(\n",
|
363 |
+
" f\"{i}\\n\"\n",
|
364 |
+
" f\"{format_timestamp(segment['start_time'], always_include_hours=True, decimal_marker=',')} --> \"\n",
|
365 |
+
" f\"{format_timestamp(segment['end_time'], always_include_hours=True, decimal_marker=',')}\\n\"\n",
|
366 |
+
" f\"{segment['speaker']}: {segment['text'].strip().replace('-->', '->')}\\n\",\n",
|
367 |
+
" file=file,\n",
|
368 |
+
" flush=True,\n",
|
369 |
+
" )\n",
|
370 |
+
"\n",
|
371 |
+
"\n",
|
372 |
+
"def find_numeral_symbol_tokens(tokenizer):\n",
|
373 |
+
" numeral_symbol_tokens = [\n",
|
374 |
+
" -1,\n",
|
375 |
+
" ]\n",
|
376 |
+
" for token, token_id in tokenizer.get_vocab().items():\n",
|
377 |
+
" has_numeral_symbol = any(c in \"0123456789%$£\" for c in token)\n",
|
378 |
+
" if has_numeral_symbol:\n",
|
379 |
+
" numeral_symbol_tokens.append(token_id)\n",
|
380 |
+
" return numeral_symbol_tokens\n",
|
381 |
+
"\n",
|
382 |
+
"\n",
|
383 |
+
"def _get_next_start_timestamp(word_timestamps, current_word_index, final_timestamp):\n",
|
384 |
+
" # if current word is the last word\n",
|
385 |
+
" if current_word_index == len(word_timestamps) - 1:\n",
|
386 |
+
" return word_timestamps[current_word_index][\"start\"]\n",
|
387 |
+
"\n",
|
388 |
+
" next_word_index = current_word_index + 1\n",
|
389 |
+
" while current_word_index < len(word_timestamps) - 1:\n",
|
390 |
+
" if word_timestamps[next_word_index].get(\"start\") is None:\n",
|
391 |
+
" # if next word doesn't have a start timestamp\n",
|
392 |
+
" # merge it with the current word and delete it\n",
|
393 |
+
" word_timestamps[current_word_index][\"word\"] += (\n",
|
394 |
+
" \" \" + word_timestamps[next_word_index][\"word\"]\n",
|
395 |
+
" )\n",
|
396 |
+
"\n",
|
397 |
+
" word_timestamps[next_word_index][\"word\"] = None\n",
|
398 |
+
" next_word_index += 1\n",
|
399 |
+
" if next_word_index == len(word_timestamps):\n",
|
400 |
+
" return final_timestamp\n",
|
401 |
+
"\n",
|
402 |
+
" else:\n",
|
403 |
+
" return word_timestamps[next_word_index][\"start\"]\n",
|
404 |
+
"\n",
|
405 |
+
"\n",
|
406 |
+
"def filter_missing_timestamps(\n",
|
407 |
+
" word_timestamps, initial_timestamp=0, final_timestamp=None\n",
|
408 |
+
"):\n",
|
409 |
+
" # handle the first and last word\n",
|
410 |
+
" if word_timestamps[0].get(\"start\") is None:\n",
|
411 |
+
" word_timestamps[0][\"start\"] = (\n",
|
412 |
+
" initial_timestamp if initial_timestamp is not None else 0\n",
|
413 |
+
" )\n",
|
414 |
+
" word_timestamps[0][\"end\"] = _get_next_start_timestamp(\n",
|
415 |
+
" word_timestamps, 0, final_timestamp\n",
|
416 |
+
" )\n",
|
417 |
+
"\n",
|
418 |
+
" result = [\n",
|
419 |
+
" word_timestamps[0],\n",
|
420 |
+
" ]\n",
|
421 |
+
"\n",
|
422 |
+
" for i, ws in enumerate(word_timestamps[1:], start=1):\n",
|
423 |
+
" # if ws doesn't have a start and end\n",
|
424 |
+
" # use the previous end as start and next start as end\n",
|
425 |
+
" if ws.get(\"start\") is None and ws.get(\"word\") is not None:\n",
|
426 |
+
" ws[\"start\"] = word_timestamps[i - 1][\"end\"]\n",
|
427 |
+
" ws[\"end\"] = _get_next_start_timestamp(word_timestamps, i, final_timestamp)\n",
|
428 |
+
"\n",
|
429 |
+
" if ws[\"word\"] is not None:\n",
|
430 |
+
" result.append(ws)\n",
|
431 |
+
" return result\n",
|
432 |
+
"\n",
|
433 |
+
"\n",
|
434 |
+
"def cleanup(path: str):\n",
|
435 |
+
" \"\"\"path could either be relative or absolute.\"\"\"\n",
|
436 |
+
" # check if file or directory exists\n",
|
437 |
+
" if os.path.isfile(path) or os.path.islink(path):\n",
|
438 |
+
" # remove file\n",
|
439 |
+
" os.remove(path)\n",
|
440 |
+
" elif os.path.isdir(path):\n",
|
441 |
+
" # remove directory and all its content\n",
|
442 |
+
" shutil.rmtree(path)\n",
|
443 |
+
" else:\n",
|
444 |
+
" raise ValueError(\"Path {} is not a file or dir.\".format(path))\n",
|
445 |
+
"\n",
|
446 |
+
"\n",
|
447 |
+
"def process_language_arg(language: str, model_name: str):\n",
|
448 |
+
" \"\"\"\n",
|
449 |
+
" Process the language argument to make sure it's valid and convert language names to language codes.\n",
|
450 |
+
" \"\"\"\n",
|
451 |
+
" if language is not None:\n",
|
452 |
+
" language = language.lower()\n",
|
453 |
+
" if language not in LANGUAGES:\n",
|
454 |
+
" if language in TO_LANGUAGE_CODE:\n",
|
455 |
+
" language = TO_LANGUAGE_CODE[language]\n",
|
456 |
+
" else:\n",
|
457 |
+
" raise ValueError(f\"Unsupported language: {language}\")\n",
|
458 |
+
"\n",
|
459 |
+
" if model_name.endswith(\".en\") and language != \"en\":\n",
|
460 |
+
" if language is not None:\n",
|
461 |
+
" logging.warning(\n",
|
462 |
+
" f\"{model_name} is an English-only model but received '{language}'; using English instead.\"\n",
|
463 |
+
" )\n",
|
464 |
+
" language = \"en\"\n",
|
465 |
+
" return language\n",
|
466 |
+
"\n",
|
467 |
+
"\n",
|
468 |
+
"def transcribe(\n",
|
469 |
+
" audio_file: str,\n",
|
470 |
+
" language: str,\n",
|
471 |
+
" model_name: str,\n",
|
472 |
+
" compute_dtype: str,\n",
|
473 |
+
" suppress_numerals: bool,\n",
|
474 |
+
" device: str,\n",
|
475 |
+
"):\n",
|
476 |
+
" from faster_whisper import WhisperModel\n",
|
477 |
+
" from helpers import find_numeral_symbol_tokens, wav2vec2_langs\n",
|
478 |
+
"\n",
|
479 |
+
" # Faster Whisper non-batched\n",
|
480 |
+
" # Run on GPU with FP16\n",
|
481 |
+
" whisper_model = WhisperModel(model_name, device=device, compute_type=compute_dtype)\n",
|
482 |
+
"\n",
|
483 |
+
" # or run on GPU with INT8\n",
|
484 |
+
" # model = WhisperModel(model_size, device=\"cuda\", compute_type=\"int8_float16\")\n",
|
485 |
+
" # or run on CPU with INT8\n",
|
486 |
+
" # model = WhisperModel(model_size, device=\"cpu\", compute_type=\"int8\")\n",
|
487 |
+
"\n",
|
488 |
+
" if suppress_numerals:\n",
|
489 |
+
" numeral_symbol_tokens = find_numeral_symbol_tokens(whisper_model.hf_tokenizer)\n",
|
490 |
+
" else:\n",
|
491 |
+
" numeral_symbol_tokens = None\n",
|
492 |
+
"\n",
|
493 |
+
" if language is not None and language in wav2vec2_langs:\n",
|
494 |
+
" word_timestamps = False\n",
|
495 |
+
" else:\n",
|
496 |
+
" word_timestamps = True\n",
|
497 |
+
"\n",
|
498 |
+
" segments, info = whisper_model.transcribe(\n",
|
499 |
+
" audio_file,\n",
|
500 |
+
" language=language,\n",
|
501 |
+
" beam_size=5,\n",
|
502 |
+
" word_timestamps=word_timestamps, # TODO: disable this if the language is supported by wav2vec2\n",
|
503 |
+
" suppress_tokens=numeral_symbol_tokens,\n",
|
504 |
+
" vad_filter=True,\n",
|
505 |
+
" )\n",
|
506 |
+
" whisper_results = []\n",
|
507 |
+
" for segment in segments:\n",
|
508 |
+
" whisper_results.append(segment._asdict())\n",
|
509 |
+
" # clear gpu vram\n",
|
510 |
+
" del whisper_model\n",
|
511 |
+
" torch.cuda.empty_cache()\n",
|
512 |
+
" return whisper_results, language\n",
|
513 |
+
"\n",
|
514 |
+
"\n",
|
515 |
+
"def transcribe_batched(\n",
|
516 |
+
" audio_file: str,\n",
|
517 |
+
" language: str,\n",
|
518 |
+
" batch_size: int,\n",
|
519 |
+
" model_name: str,\n",
|
520 |
+
" compute_dtype: str,\n",
|
521 |
+
" suppress_numerals: bool,\n",
|
522 |
+
" device: str,\n",
|
523 |
+
"):\n",
|
524 |
+
" import whisperx\n",
|
525 |
+
"\n",
|
526 |
+
" # Faster Whisper batched\n",
|
527 |
+
" whisper_model = whisperx.load_model(\n",
|
528 |
+
" model_name,\n",
|
529 |
+
" device,\n",
|
530 |
+
" compute_type=compute_dtype,\n",
|
531 |
+
" asr_options={\"suppress_numerals\": suppress_numerals},\n",
|
532 |
+
" )\n",
|
533 |
+
" audio = whisperx.load_audio(audio_file)\n",
|
534 |
+
" result = whisper_model.transcribe(audio, language=language, batch_size=batch_size)\n",
|
535 |
+
" del whisper_model\n",
|
536 |
+
" torch.cuda.empty_cache()\n",
|
537 |
+
" return result[\"segments\"], result[\"language\"]"
|
538 |
+
]
|
539 |
+
},
|
540 |
+
{
|
541 |
+
"attachments": {},
|
542 |
+
"cell_type": "markdown",
|
543 |
+
"metadata": {
|
544 |
+
"id": "B7qWQb--1Xcw"
|
545 |
+
},
|
546 |
+
"source": [
|
547 |
+
"# Options"
|
548 |
+
]
|
549 |
+
},
|
550 |
+
{
|
551 |
+
"cell_type": "code",
|
552 |
+
"execution_count": null,
|
553 |
+
"metadata": {
|
554 |
+
"id": "ONlFrSnD0FOp"
|
555 |
+
},
|
556 |
+
"outputs": [],
|
557 |
+
"source": [
|
558 |
+
"# Name of the audio file\n",
|
559 |
+
"audio_path = \"20200128-Pieter Wuille (part 1 of 2) - Episode 1.mp3\"\n",
|
560 |
+
"\n",
|
561 |
+
"# Whether to enable music removal from speech, helps increase diarization quality but uses alot of ram\n",
|
562 |
+
"enable_stemming = True\n",
|
563 |
+
"\n",
|
564 |
+
"# (choose from 'tiny.en', 'tiny', 'base.en', 'base', 'small.en', 'small', 'medium.en', 'medium', 'large-v1', 'large-v2', 'large-v3', 'large')\n",
|
565 |
+
"whisper_model_name = \"large-v2\"\n",
|
566 |
+
"\n",
|
567 |
+
"# replaces numerical digits with their pronounciation, increases diarization accuracy\n",
|
568 |
+
"suppress_numerals = True\n",
|
569 |
+
"\n",
|
570 |
+
"batch_size = 8\n",
|
571 |
+
"\n",
|
572 |
+
"language = None # autodetect language\n",
|
573 |
+
"\n",
|
574 |
+
"device = \"cuda\" if torch.cuda.is_available() else \"cpu\""
|
575 |
+
]
|
576 |
+
},
|
577 |
+
{
|
578 |
+
"attachments": {},
|
579 |
+
"cell_type": "markdown",
|
580 |
+
"metadata": {
|
581 |
+
"id": "h-cY1ZEy2KVI"
|
582 |
+
},
|
583 |
+
"source": [
|
584 |
+
"# Processing"
|
585 |
+
]
|
586 |
+
},
|
587 |
+
{
|
588 |
+
"attachments": {},
|
589 |
+
"cell_type": "markdown",
|
590 |
+
"metadata": {
|
591 |
+
"id": "7ZS4xXmE2NGP"
|
592 |
+
},
|
593 |
+
"source": [
|
594 |
+
"## Separating music from speech using Demucs\n",
|
595 |
+
"\n",
|
596 |
+
"---\n",
|
597 |
+
"\n",
|
598 |
+
"By isolating the vocals from the rest of the audio, it becomes easier to identify and track individual speakers based on the spectral and temporal characteristics of their speech signals. Source separation is just one of many techniques that can be used as a preprocessing step to help improve the accuracy and reliability of the overall diarization process."
|
599 |
+
]
|
600 |
+
},
|
601 |
+
{
|
602 |
+
"cell_type": "code",
|
603 |
+
"execution_count": null,
|
604 |
+
"metadata": {
|
605 |
+
"colab": {
|
606 |
+
"base_uri": "https://localhost:8080/"
|
607 |
+
},
|
608 |
+
"id": "HKcgQUrAzsJZ",
|
609 |
+
"outputId": "dc2a1d96-20da-4749-9d64-21edacfba1b1"
|
610 |
+
},
|
611 |
+
"outputs": [],
|
612 |
+
"source": [
|
613 |
+
"if enable_stemming:\n",
|
614 |
+
" # Isolate vocals from the rest of the audio\n",
|
615 |
+
"\n",
|
616 |
+
" return_code = os.system(\n",
|
617 |
+
" f'python3 -m demucs.separate -n htdemucs --two-stems=vocals \"{audio_path}\" -o \"temp_outputs\"'\n",
|
618 |
+
" )\n",
|
619 |
+
"\n",
|
620 |
+
" if return_code != 0:\n",
|
621 |
+
" logging.warning(\"Source splitting failed, using original audio file.\")\n",
|
622 |
+
" vocal_target = audio_path\n",
|
623 |
+
" else:\n",
|
624 |
+
" vocal_target = os.path.join(\n",
|
625 |
+
" \"temp_outputs\",\n",
|
626 |
+
" \"htdemucs\",\n",
|
627 |
+
" os.path.splitext(os.path.basename(audio_path))[0],\n",
|
628 |
+
" \"vocals.wav\",\n",
|
629 |
+
" )\n",
|
630 |
+
"else:\n",
|
631 |
+
" vocal_target = audio_path"
|
632 |
+
]
|
633 |
+
},
|
634 |
+
{
|
635 |
+
"attachments": {},
|
636 |
+
"cell_type": "markdown",
|
637 |
+
"metadata": {
|
638 |
+
"id": "UYg9VWb22Tz8"
|
639 |
+
},
|
640 |
+
"source": [
|
641 |
+
"## Transcriping audio using Whisper and realligning timestamps using Wav2Vec2\n",
|
642 |
+
"---\n",
|
643 |
+
"This code uses two different open-source models to transcribe speech and perform forced alignment on the resulting transcription.\n",
|
644 |
+
"\n",
|
645 |
+
"The first model is called OpenAI Whisper, which is a speech recognition model that can transcribe speech with high accuracy. The code loads the whisper model and uses it to transcribe the vocal_target file.\n",
|
646 |
+
"\n",
|
647 |
+
"The output of the transcription process is a set of text segments with corresponding timestamps indicating when each segment was spoken.\n"
|
648 |
+
]
|
649 |
+
},
|
650 |
+
{
|
651 |
+
"cell_type": "code",
|
652 |
+
"execution_count": null,
|
653 |
+
"metadata": {
|
654 |
+
"id": "5-VKFn530oTl"
|
655 |
+
},
|
656 |
+
"outputs": [],
|
657 |
+
"source": [
|
658 |
+
"compute_type = \"float16\"\n",
|
659 |
+
"# or run on GPU with INT8\n",
|
660 |
+
"# compute_type = \"int8_float16\"\n",
|
661 |
+
"# or run on CPU with INT8\n",
|
662 |
+
"# compute_type = \"int8\"\n",
|
663 |
+
"\n",
|
664 |
+
"if batch_size != 0:\n",
|
665 |
+
" whisper_results, language = transcribe_batched(\n",
|
666 |
+
" vocal_target,\n",
|
667 |
+
" language,\n",
|
668 |
+
" batch_size,\n",
|
669 |
+
" whisper_model_name,\n",
|
670 |
+
" compute_type,\n",
|
671 |
+
" suppress_numerals,\n",
|
672 |
+
" device,\n",
|
673 |
+
" )\n",
|
674 |
+
"else:\n",
|
675 |
+
" whisper_results, language = transcribe(\n",
|
676 |
+
" vocal_target,\n",
|
677 |
+
" language,\n",
|
678 |
+
" whisper_model_name,\n",
|
679 |
+
" compute_type,\n",
|
680 |
+
" suppress_numerals,\n",
|
681 |
+
" device,\n",
|
682 |
+
" )"
|
683 |
+
]
|
684 |
+
},
|
685 |
+
{
|
686 |
+
"attachments": {},
|
687 |
+
"cell_type": "markdown",
|
688 |
+
"metadata": {},
|
689 |
+
"source": [
|
690 |
+
"## Aligning the transcription with the original audio using Wav2Vec2\n",
|
691 |
+
"---\n",
|
692 |
+
"The second model used is called wav2vec2, which is a large-scale neural network that is designed to learn representations of speech that are useful for a variety of speech processing tasks, including speech recognition and alignment.\n",
|
693 |
+
"\n",
|
694 |
+
"The code loads the wav2vec2 alignment model and uses it to align the transcription segments with the original audio signal contained in the vocal_target file. This process involves finding the exact timestamps in the audio signal where each segment was spoken and aligning the text accordingly.\n",
|
695 |
+
"\n",
|
696 |
+
"By combining the outputs of the two models, the code produces a fully aligned transcription of the speech contained in the vocal_target file. This aligned transcription can be useful for a variety of speech processing tasks, such as speaker diarization, sentiment analysis, and language identification.\n",
|
697 |
+
"\n",
|
698 |
+
"If there's no Wav2Vec2 model available for your language, word timestamps generated by whisper will be used instead."
|
699 |
+
]
|
700 |
+
},
|
701 |
+
{
|
702 |
+
"cell_type": "code",
|
703 |
+
"execution_count": null,
|
704 |
+
"metadata": {},
|
705 |
+
"outputs": [],
|
706 |
+
"source": [
|
707 |
+
"if language in wav2vec2_langs:\n",
|
708 |
+
" device = \"cuda\"\n",
|
709 |
+
" alignment_model, metadata = whisperx.load_align_model(\n",
|
710 |
+
" language_code=language, device=device\n",
|
711 |
+
" )\n",
|
712 |
+
" result_aligned = whisperx.align(\n",
|
713 |
+
" whisper_results, alignment_model, metadata, vocal_target, device\n",
|
714 |
+
" )\n",
|
715 |
+
" word_timestamps = filter_missing_timestamps(\n",
|
716 |
+
" result_aligned[\"word_segments\"],\n",
|
717 |
+
" initial_timestamp=whisper_results[0].get(\"start\"),\n",
|
718 |
+
" final_timestamp=whisper_results[-1].get(\"end\"),\n",
|
719 |
+
" )\n",
|
720 |
+
"\n",
|
721 |
+
" # clear gpu vram\n",
|
722 |
+
" del alignment_model\n",
|
723 |
+
" torch.cuda.empty_cache()\n",
|
724 |
+
"else:\n",
|
725 |
+
" assert batch_size == 0, ( # TODO: add a better check for word timestamps existence\n",
|
726 |
+
" f\"Unsupported language: {language}, use --batch_size to 0\"\n",
|
727 |
+
" \" to generate word timestamps using whisper directly and fix this error.\"\n",
|
728 |
+
" )\n",
|
729 |
+
" word_timestamps = []\n",
|
730 |
+
" for segment in whisper_results:\n",
|
731 |
+
" for word in segment[\"words\"]:\n",
|
732 |
+
" word_timestamps.append({\"word\": word[2], \"start\": word[0], \"end\": word[1]})"
|
733 |
+
]
|
734 |
+
},
|
735 |
+
{
|
736 |
+
"attachments": {},
|
737 |
+
"cell_type": "markdown",
|
738 |
+
"metadata": {
|
739 |
+
"id": "7EEaJPsQ21Rx"
|
740 |
+
},
|
741 |
+
"source": [
|
742 |
+
"## Convert audio to mono for NeMo combatibility"
|
743 |
+
]
|
744 |
+
},
|
745 |
+
{
|
746 |
+
"cell_type": "code",
|
747 |
+
"execution_count": null,
|
748 |
+
"metadata": {},
|
749 |
+
"outputs": [],
|
750 |
+
"source": [
|
751 |
+
"sound = AudioSegment.from_file(vocal_target).set_channels(1)\n",
|
752 |
+
"ROOT = os.getcwd()\n",
|
753 |
+
"temp_path = os.path.join(ROOT, \"temp_outputs\")\n",
|
754 |
+
"os.makedirs(temp_path, exist_ok=True)\n",
|
755 |
+
"sound.export(os.path.join(temp_path, \"mono_file.wav\"), format=\"wav\")"
|
756 |
+
]
|
757 |
+
},
|
758 |
+
{
|
759 |
+
"attachments": {},
|
760 |
+
"cell_type": "markdown",
|
761 |
+
"metadata": {
|
762 |
+
"id": "D1gkViCf2-CV"
|
763 |
+
},
|
764 |
+
"source": [
|
765 |
+
"## Speaker Diarization using NeMo MSDD Model\n",
|
766 |
+
"---\n",
|
767 |
+
"This code uses a model called Nvidia NeMo MSDD (Multi-scale Diarization Decoder) to perform speaker diarization on an audio signal. Speaker diarization is the process of separating an audio signal into different segments based on who is speaking at any given time."
|
768 |
+
]
|
769 |
+
},
|
770 |
+
{
|
771 |
+
"cell_type": "code",
|
772 |
+
"execution_count": null,
|
773 |
+
"metadata": {
|
774 |
+
"id": "C7jIpBCH02RL"
|
775 |
+
},
|
776 |
+
"outputs": [],
|
777 |
+
"source": [
|
778 |
+
"# Initialize NeMo MSDD diarization model\n",
|
779 |
+
"msdd_model = NeuralDiarizer(cfg=create_config(temp_path)).to(\"cuda\")\n",
|
780 |
+
"msdd_model.diarize()\n",
|
781 |
+
"\n",
|
782 |
+
"del msdd_model\n",
|
783 |
+
"torch.cuda.empty_cache()"
|
784 |
+
]
|
785 |
+
},
|
786 |
+
{
|
787 |
+
"attachments": {},
|
788 |
+
"cell_type": "markdown",
|
789 |
+
"metadata": {
|
790 |
+
"id": "NmkZYaDAEOAg"
|
791 |
+
},
|
792 |
+
"source": [
|
793 |
+
"## Mapping Spekers to Sentences According to Timestamps"
|
794 |
+
]
|
795 |
+
},
|
796 |
+
{
|
797 |
+
"cell_type": "code",
|
798 |
+
"execution_count": null,
|
799 |
+
"metadata": {
|
800 |
+
"id": "E65LUGQe02zw"
|
801 |
+
},
|
802 |
+
"outputs": [],
|
803 |
+
"source": [
|
804 |
+
"# Reading timestamps <> Speaker Labels mapping\n",
|
805 |
+
"\n",
|
806 |
+
"speaker_ts = []\n",
|
807 |
+
"with open(os.path.join(temp_path, \"pred_rttms\", \"mono_file.rttm\"), \"r\") as f:\n",
|
808 |
+
" lines = f.readlines()\n",
|
809 |
+
" for line in lines:\n",
|
810 |
+
" line_list = line.split(\" \")\n",
|
811 |
+
" s = int(float(line_list[5]) * 1000)\n",
|
812 |
+
" e = s + int(float(line_list[8]) * 1000)\n",
|
813 |
+
" speaker_ts.append([s, e, int(line_list[11].split(\"_\")[-1])])\n",
|
814 |
+
"\n",
|
815 |
+
"wsm = get_words_speaker_mapping(word_timestamps, speaker_ts, \"start\")"
|
816 |
+
]
|
817 |
+
},
|
818 |
+
{
|
819 |
+
"attachments": {},
|
820 |
+
"cell_type": "markdown",
|
821 |
+
"metadata": {
|
822 |
+
"id": "8Ruxc8S1EXtW"
|
823 |
+
},
|
824 |
+
"source": [
|
825 |
+
"## Realligning Speech segments using Punctuation\n",
|
826 |
+
"---\n",
|
827 |
+
"\n",
|
828 |
+
"This code provides a method for disambiguating speaker labels in cases where a sentence is split between two different speakers. It uses punctuation markings to determine the dominant speaker for each sentence in the transcription.\n",
|
829 |
+
"\n",
|
830 |
+
"```\n",
|
831 |
+
"Speaker A: It's got to come from somewhere else. Yeah, that one's also fun because you know the lows are\n",
|
832 |
+
"Speaker B: going to suck, right? So it's actually it hits you on both sides.\n",
|
833 |
+
"```\n",
|
834 |
+
"\n",
|
835 |
+
"For example, if a sentence is split between two speakers, the code takes the mode of speaker labels for each word in the sentence, and uses that speaker label for the whole sentence. This can help to improve the accuracy of speaker diarization, especially in cases where the Whisper model may not take fine utterances like \"hmm\" and \"yeah\" into account, but the Diarization Model (Nemo) may include them, leading to inconsistent results.\n",
|
836 |
+
"\n",
|
837 |
+
"The code also handles cases where one speaker is giving a monologue while other speakers are making occasional comments in the background. It ignores the comments and assigns the entire monologue to the speaker who is speaking the majority of the time. This provides a robust and reliable method for realigning speech segments to their respective speakers based on punctuation in the transcription."
|
838 |
+
]
|
839 |
+
},
|
840 |
+
{
|
841 |
+
"cell_type": "code",
|
842 |
+
"execution_count": null,
|
843 |
+
"metadata": {
|
844 |
+
"id": "pgfC5hA41BXu"
|
845 |
+
},
|
846 |
+
"outputs": [],
|
847 |
+
"source": [
|
848 |
+
"if language in punct_model_langs:\n",
|
849 |
+
" # restoring punctuation in the transcript to help realign the sentences\n",
|
850 |
+
" punct_model = PunctuationModel(model=\"kredor/punctuate-all\")\n",
|
851 |
+
"\n",
|
852 |
+
" words_list = list(map(lambda x: x[\"word\"], wsm))\n",
|
853 |
+
"\n",
|
854 |
+
" labled_words = punct_model.predict(words_list)\n",
|
855 |
+
"\n",
|
856 |
+
" ending_puncts = \".?!\"\n",
|
857 |
+
" model_puncts = \".,;:!?\"\n",
|
858 |
+
"\n",
|
859 |
+
" # We don't want to punctuate U.S.A. with a period. Right?\n",
|
860 |
+
" is_acronym = lambda x: re.fullmatch(r\"\\b(?:[a-zA-Z]\\.){2,}\", x)\n",
|
861 |
+
"\n",
|
862 |
+
" for word_dict, labeled_tuple in zip(wsm, labled_words):\n",
|
863 |
+
" word = word_dict[\"word\"]\n",
|
864 |
+
" if (\n",
|
865 |
+
" word\n",
|
866 |
+
" and labeled_tuple[1] in ending_puncts\n",
|
867 |
+
" and (word[-1] not in model_puncts or is_acronym(word))\n",
|
868 |
+
" ):\n",
|
869 |
+
" word += labeled_tuple[1]\n",
|
870 |
+
" if word.endswith(\"..\"):\n",
|
871 |
+
" word = word.rstrip(\".\")\n",
|
872 |
+
" word_dict[\"word\"] = word\n",
|
873 |
+
"\n",
|
874 |
+
"else:\n",
|
875 |
+
" logging.warning(\n",
|
876 |
+
" f\"Punctuation restoration is not available for {language} language. Using the original punctuation.\"\n",
|
877 |
+
" )\n",
|
878 |
+
"\n",
|
879 |
+
"wsm = get_realigned_ws_mapping_with_punctuation(wsm)\n",
|
880 |
+
"ssm = get_sentences_speaker_mapping(wsm, speaker_ts)"
|
881 |
+
]
|
882 |
+
},
|
883 |
+
{
|
884 |
+
"attachments": {},
|
885 |
+
"cell_type": "markdown",
|
886 |
+
"metadata": {
|
887 |
+
"id": "vF2QAtLOFvwZ"
|
888 |
+
},
|
889 |
+
"source": [
|
890 |
+
"## Cleanup and Exporing the results"
|
891 |
+
]
|
892 |
+
},
|
893 |
+
{
|
894 |
+
"cell_type": "code",
|
895 |
+
"execution_count": null,
|
896 |
+
"metadata": {
|
897 |
+
"id": "kFTyKI6B1MI0"
|
898 |
+
},
|
899 |
+
"outputs": [],
|
900 |
+
"source": [
|
901 |
+
"with open(f\"{os.path.splitext(audio_path)[0]}.txt\", \"w\", encoding=\"utf-8-sig\") as f:\n",
|
902 |
+
" get_speaker_aware_transcript(ssm, f)\n",
|
903 |
+
"\n",
|
904 |
+
"with open(f\"{os.path.splitext(audio_path)[0]}.srt\", \"w\", encoding=\"utf-8-sig\") as srt:\n",
|
905 |
+
" write_srt(ssm, srt)\n",
|
906 |
+
"\n",
|
907 |
+
"cleanup(temp_path)"
|
908 |
+
]
|
909 |
+
}
|
910 |
+
],
|
911 |
+
"metadata": {
|
912 |
+
"accelerator": "GPU",
|
913 |
+
"colab": {
|
914 |
+
"authorship_tag": "ABX9TyOyiQNkD+ROzss634BOsrSh",
|
915 |
+
"collapsed_sections": [
|
916 |
+
"eCmjcOc9yEtQ",
|
917 |
+
"jbsUt3SwyhjD"
|
918 |
+
],
|
919 |
+
"include_colab_link": true,
|
920 |
+
"provenance": []
|
921 |
+
},
|
922 |
+
"gpuClass": "standard",
|
923 |
+
"kernelspec": {
|
924 |
+
"display_name": "Python 3",
|
925 |
+
"name": "python3"
|
926 |
+
},
|
927 |
+
"language_info": {
|
928 |
+
"codemirror_mode": {
|
929 |
+
"name": "ipython",
|
930 |
+
"version": 3
|
931 |
+
},
|
932 |
+
"file_extension": ".py",
|
933 |
+
"mimetype": "text/x-python",
|
934 |
+
"name": "python",
|
935 |
+
"nbconvert_exporter": "python",
|
936 |
+
"pygments_lexer": "ipython3",
|
937 |
+
"version": "3.10.12"
|
938 |
+
}
|
939 |
+
},
|
940 |
+
"nbformat": 4,
|
941 |
+
"nbformat_minor": 0
|
942 |
+
}
|
audio_clean.py
ADDED
@@ -0,0 +1,126 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import numpy as np
|
3 |
+
import librosa
|
4 |
+
import soundfile as sf
|
5 |
+
from pydub import AudioSegment
|
6 |
+
from pydub.silence import split_on_silence
|
7 |
+
from pydub.playback import play
|
8 |
+
from tqdm import tqdm
|
9 |
+
|
10 |
+
def clean_audio(audio_path, output_path, selected_chunks, min_silence_len=1000, silence_thresh=-40, keep_silence=100):
|
11 |
+
# Load the audio file
|
12 |
+
audio_segment = AudioSegment.from_file(audio_path)
|
13 |
+
|
14 |
+
# Convert to mono
|
15 |
+
audio_segment = audio_segment.set_channels(1)
|
16 |
+
|
17 |
+
# Normalize the audio
|
18 |
+
audio_segment = normalize_audio(audio_segment)
|
19 |
+
|
20 |
+
# Split on silence
|
21 |
+
chunks = split_on_silence(
|
22 |
+
audio_segment,
|
23 |
+
min_silence_len=min_silence_len,
|
24 |
+
silence_thresh=silence_thresh,
|
25 |
+
keep_silence=keep_silence,
|
26 |
+
)
|
27 |
+
|
28 |
+
# Find the main speaker based on total duration
|
29 |
+
main_speaker_chunk = max(chunks, key=lambda chunk: len(chunk))
|
30 |
+
|
31 |
+
# Apply EQ and compression
|
32 |
+
main_speaker_chunk = apply_eq_and_compression(main_speaker_chunk)
|
33 |
+
|
34 |
+
# Export the main speaker's audio
|
35 |
+
main_speaker_chunk.export(output_path, format="wav")
|
36 |
+
|
37 |
+
def normalize_audio(audio_segment):
|
38 |
+
"""
|
39 |
+
Normalizes the audio to a target volume.
|
40 |
+
"""
|
41 |
+
target_dBFS = -20
|
42 |
+
change_in_dBFS = target_dBFS - audio_segment.dBFS
|
43 |
+
return audio_segment.apply_gain(change_in_dBFS)
|
44 |
+
|
45 |
+
def apply_eq_and_compression(audio_segment):
|
46 |
+
"""
|
47 |
+
Applies equalization and compression to the audio.
|
48 |
+
"""
|
49 |
+
# Apply EQ
|
50 |
+
audio_segment = audio_segment.high_pass_filter(80)
|
51 |
+
audio_segment = audio_segment.low_pass_filter(12000)
|
52 |
+
|
53 |
+
# Apply compression
|
54 |
+
threshold = -20
|
55 |
+
ratio = 2
|
56 |
+
attack = 10
|
57 |
+
release = 100
|
58 |
+
audio_segment = audio_segment.compress_dynamic_range(
|
59 |
+
threshold=threshold,
|
60 |
+
ratio=ratio,
|
61 |
+
attack=attack,
|
62 |
+
release=release,
|
63 |
+
)
|
64 |
+
|
65 |
+
return audio_segment
|
66 |
+
|
67 |
+
def process_file(wav_file, srt_file, cleaned_folder):
|
68 |
+
print(f"Processing file: {wav_file}")
|
69 |
+
|
70 |
+
# Create the cleaned folder if it doesn't exist
|
71 |
+
os.makedirs(cleaned_folder, exist_ok=True)
|
72 |
+
|
73 |
+
input_wav_path = wav_file
|
74 |
+
output_wav_path = os.path.join(cleaned_folder, os.path.basename(wav_file))
|
75 |
+
|
76 |
+
# Review and select desired SRT chunks
|
77 |
+
selected_chunks = review_srt_chunks(input_wav_path, srt_file)
|
78 |
+
|
79 |
+
# Clean the audio based on selected chunks
|
80 |
+
clean_audio(input_wav_path, output_wav_path, selected_chunks)
|
81 |
+
|
82 |
+
print(f"Cleaned audio saved to: {output_wav_path}")
|
83 |
+
|
84 |
+
def review_srt_chunks(audio_path, srt_path):
|
85 |
+
audio_segment = AudioSegment.from_wav(audio_path)
|
86 |
+
selected_chunks = []
|
87 |
+
|
88 |
+
with open(srt_path, "r") as srt_file:
|
89 |
+
srt_content = srt_file.read()
|
90 |
+
srt_entries = srt_content.strip().split("\n\n")
|
91 |
+
|
92 |
+
for entry in tqdm(srt_entries, desc="Reviewing SRT chunks", unit="chunk"):
|
93 |
+
lines = entry.strip().split("\n")
|
94 |
+
if len(lines) >= 3:
|
95 |
+
start_time, end_time = lines[1].split(" --> ")
|
96 |
+
start_time = convert_to_milliseconds(start_time)
|
97 |
+
end_time = convert_to_milliseconds(end_time)
|
98 |
+
|
99 |
+
chunk = audio_segment[start_time:end_time]
|
100 |
+
print("Playing chunk...")
|
101 |
+
play(chunk)
|
102 |
+
|
103 |
+
choice = input("Keep this chunk? (y/n): ")
|
104 |
+
if choice.lower() == "y":
|
105 |
+
selected_chunks.append((start_time, end_time))
|
106 |
+
print("Chunk selected.")
|
107 |
+
else:
|
108 |
+
print("Chunk skipped.")
|
109 |
+
|
110 |
+
return selected_chunks
|
111 |
+
|
112 |
+
def convert_to_milliseconds(time_str):
|
113 |
+
time_str = time_str.replace(",", ".")
|
114 |
+
hours, minutes, seconds = time_str.strip().split(":")
|
115 |
+
milliseconds = (int(hours) * 3600 + int(minutes) * 60 + float(seconds)) * 1000
|
116 |
+
return int(milliseconds)
|
117 |
+
|
118 |
+
# Set the WAV file, SRT file, and cleaned folder paths
|
119 |
+
wav_file = "/path/to/your/audio.wav"
|
120 |
+
srt_file = "/path/to/your/subtitles.srt"
|
121 |
+
cleaned_folder = "/path/to/cleaned/folder"
|
122 |
+
|
123 |
+
# Process the WAV file
|
124 |
+
process_file(wav_file, srt_file, cleaned_folder)
|
125 |
+
|
126 |
+
print("Processing completed.")
|
audio_cleaning_test.py
ADDED
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import librosa
|
3 |
+
import numpy as np
|
4 |
+
from pydub import AudioSegment
|
5 |
+
|
6 |
+
def clean_audio(audio_path, output_path, min_silence_len=1000, silence_thresh=-40, keep_silence=100):
|
7 |
+
# Load the audio file
|
8 |
+
audio_segment = AudioSegment.from_file(audio_path)
|
9 |
+
|
10 |
+
# Convert to mono
|
11 |
+
audio_segment = audio_segment.set_channels(1)
|
12 |
+
|
13 |
+
# Split on silence
|
14 |
+
chunks = split_on_silence(
|
15 |
+
audio_segment,
|
16 |
+
min_silence_len=min_silence_len,
|
17 |
+
silence_thresh=silence_thresh,
|
18 |
+
keep_silence=keep_silence,
|
19 |
+
)
|
20 |
+
|
21 |
+
# Find the main speaker based on total duration
|
22 |
+
main_speaker_chunk = max(chunks, key=lambda chunk: len(chunk))
|
23 |
+
|
24 |
+
# Export the main speaker's audio
|
25 |
+
main_speaker_chunk.export(output_path, format="wav")
|
26 |
+
|
27 |
+
def split_on_silence(audio_segment, min_silence_len=1000, silence_thresh=-40, keep_silence=100):
|
28 |
+
"""
|
29 |
+
Splits an AudioSegment on silent sections.
|
30 |
+
"""
|
31 |
+
chunks = []
|
32 |
+
start_idx = 0
|
33 |
+
|
34 |
+
while start_idx < len(audio_segment):
|
35 |
+
silence_start = audio_segment.find_silence(
|
36 |
+
min_silence_len=min_silence_len,
|
37 |
+
silence_thresh=silence_thresh,
|
38 |
+
start_sec=start_idx / 1000.0,
|
39 |
+
)
|
40 |
+
|
41 |
+
if silence_start is None:
|
42 |
+
chunks.append(audio_segment[start_idx:])
|
43 |
+
break
|
44 |
+
|
45 |
+
silence_end = silence_start + min_silence_len
|
46 |
+
keep_silence_time = min(keep_silence, silence_end - silence_start)
|
47 |
+
silence_end -= keep_silence_time
|
48 |
+
|
49 |
+
chunks.append(audio_segment[start_idx:silence_end])
|
50 |
+
start_idx = silence_end + keep_silence_time
|
51 |
+
|
52 |
+
return chunks
|
53 |
+
|
54 |
+
# Usage example
|
55 |
+
audio_path = "francine-master.wav"
|
56 |
+
output_path = "franclean-master.wav"
|
57 |
+
clean_audio(audio_path, output_path)
|
autodiarize.py
ADDED
@@ -0,0 +1,255 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
from helpers import *
|
4 |
+
from faster_whisper import WhisperModel
|
5 |
+
import whisperx
|
6 |
+
import torch
|
7 |
+
from pydub import AudioSegment
|
8 |
+
from nemo.collections.asr.models.msdd_models import NeuralDiarizer
|
9 |
+
from deepmultilingualpunctuation import PunctuationModel
|
10 |
+
import re
|
11 |
+
import logging
|
12 |
+
import pysrt
|
13 |
+
|
14 |
+
mtypes = {"cpu": "int8", "cuda": "float16"}
|
15 |
+
|
16 |
+
# Initialize parser
|
17 |
+
parser = argparse.ArgumentParser()
|
18 |
+
parser.add_argument(
|
19 |
+
"-a", "--audio", help="name of the target audio file", required=True
|
20 |
+
)
|
21 |
+
parser.add_argument(
|
22 |
+
"-s", "--srt", help="name of the target SRT file", required=True
|
23 |
+
)
|
24 |
+
parser.add_argument(
|
25 |
+
"--no-stem",
|
26 |
+
action="store_false",
|
27 |
+
dest="stemming",
|
28 |
+
default=True,
|
29 |
+
help="Disables source separation."
|
30 |
+
"This helps with long files that don't contain a lot of music.",
|
31 |
+
)
|
32 |
+
parser.add_argument(
|
33 |
+
"--suppress_numerals",
|
34 |
+
action="store_true",
|
35 |
+
dest="suppress_numerals",
|
36 |
+
default=False,
|
37 |
+
help="Suppresses Numerical Digits."
|
38 |
+
"This helps the diarization accuracy but converts all digits into written text.",
|
39 |
+
)
|
40 |
+
parser.add_argument(
|
41 |
+
"--whisper-model",
|
42 |
+
dest="model_name",
|
43 |
+
default="medium.en",
|
44 |
+
help="name of the Whisper model to use",
|
45 |
+
)
|
46 |
+
parser.add_argument(
|
47 |
+
"--batch-size",
|
48 |
+
type=int,
|
49 |
+
dest="batch_size",
|
50 |
+
default=8,
|
51 |
+
help="Batch size for batched inference, reduce if you run out of memory, set to 0 for non-batched inference",
|
52 |
+
)
|
53 |
+
parser.add_argument(
|
54 |
+
"--language",
|
55 |
+
type=str,
|
56 |
+
default=None,
|
57 |
+
choices=whisper_langs,
|
58 |
+
help="Language spoken in the audio, specify None to perform language detection",
|
59 |
+
)
|
60 |
+
parser.add_argument(
|
61 |
+
"--device",
|
62 |
+
dest="device",
|
63 |
+
default="cuda" if torch.cuda.is_available() else "cpu",
|
64 |
+
help="if you have a GPU use 'cuda', otherwise 'cpu'",
|
65 |
+
)
|
66 |
+
args = parser.parse_args()
|
67 |
+
|
68 |
+
def ensure_dir(directory):
|
69 |
+
if not os.path.exists(directory):
|
70 |
+
os.makedirs(directory)
|
71 |
+
|
72 |
+
if args.stemming:
|
73 |
+
# Isolate vocals from the rest of the audio
|
74 |
+
return_code = os.system(
|
75 |
+
f'python3 -m demucs.separate -n htdemucs --two-stems=vocals "{args.audio}" -o "temp_outputs"'
|
76 |
+
)
|
77 |
+
if return_code != 0:
|
78 |
+
logging.warning(
|
79 |
+
"Source splitting failed, using original audio file. Use --no-stem argument to disable it."
|
80 |
+
)
|
81 |
+
vocal_target = args.audio
|
82 |
+
else:
|
83 |
+
vocal_target = os.path.join(
|
84 |
+
"temp_outputs",
|
85 |
+
"htdemucs",
|
86 |
+
os.path.splitext(os.path.basename(args.audio))[0],
|
87 |
+
"vocals.wav",
|
88 |
+
)
|
89 |
+
else:
|
90 |
+
vocal_target = args.audio
|
91 |
+
|
92 |
+
# Transcribe the audio file
|
93 |
+
if args.batch_size != 0:
|
94 |
+
from transcription_helpers import transcribe_batched
|
95 |
+
whisper_results, language = transcribe_batched(
|
96 |
+
vocal_target,
|
97 |
+
args.language,
|
98 |
+
args.batch_size,
|
99 |
+
args.model_name,
|
100 |
+
mtypes[args.device],
|
101 |
+
args.suppress_numerals,
|
102 |
+
args.device,
|
103 |
+
)
|
104 |
+
else:
|
105 |
+
from transcription_helpers import transcribe
|
106 |
+
whisper_results, language = transcribe(
|
107 |
+
vocal_target,
|
108 |
+
args.language,
|
109 |
+
args.model_name,
|
110 |
+
mtypes[args.device],
|
111 |
+
args.suppress_numerals,
|
112 |
+
args.device,
|
113 |
+
)
|
114 |
+
|
115 |
+
if language in wav2vec2_langs:
|
116 |
+
alignment_model, metadata = whisperx.load_align_model(
|
117 |
+
language_code=language, device=args.device
|
118 |
+
)
|
119 |
+
result_aligned = whisperx.align(
|
120 |
+
whisper_results, alignment_model, metadata, vocal_target, args.device
|
121 |
+
)
|
122 |
+
word_timestamps = filter_missing_timestamps(
|
123 |
+
result_aligned["word_segments"],
|
124 |
+
initial_timestamp=whisper_results[0].get("start"),
|
125 |
+
final_timestamp=whisper_results[-1].get("end"),
|
126 |
+
)
|
127 |
+
# clear gpu vram
|
128 |
+
del alignment_model
|
129 |
+
torch.cuda.empty_cache()
|
130 |
+
else:
|
131 |
+
assert (
|
132 |
+
args.batch_size == 0 # TODO: add a better check for word timestamps existence
|
133 |
+
), (
|
134 |
+
f"Unsupported language: {language}, use --batch_size to 0"
|
135 |
+
" to generate word timestamps using whisper directly and fix this error."
|
136 |
+
)
|
137 |
+
word_timestamps = []
|
138 |
+
for segment in whisper_results:
|
139 |
+
for word in segment["words"]:
|
140 |
+
word_timestamps.append({"word": word[2], "start": word[0], "end": word[1]})
|
141 |
+
|
142 |
+
# convert audio to mono for NeMo compatibility
|
143 |
+
sound = AudioSegment.from_file(vocal_target).set_channels(1)
|
144 |
+
ROOT = os.getcwd()
|
145 |
+
temp_path = os.path.join(ROOT, "temp_outputs")
|
146 |
+
os.makedirs(temp_path, exist_ok=True)
|
147 |
+
sound.export(os.path.join(temp_path, "mono_file.wav"), format="wav")
|
148 |
+
|
149 |
+
# Initialize NeMo MSDD diarization model
|
150 |
+
msdd_model = NeuralDiarizer(cfg=create_config(temp_path)).to(args.device)
|
151 |
+
msdd_model.diarize()
|
152 |
+
del msdd_model
|
153 |
+
torch.cuda.empty_cache()
|
154 |
+
|
155 |
+
# Reading timestamps <> Speaker Labels mapping
|
156 |
+
speaker_ts = []
|
157 |
+
with open(os.path.join(temp_path, "pred_rttms", "mono_file.rttm"), "r") as f:
|
158 |
+
lines = f.readlines()
|
159 |
+
for line in lines:
|
160 |
+
line_list = line.split(" ")
|
161 |
+
s = int(float(line_list[5]) * 1000)
|
162 |
+
e = s + int(float(line_list[8]) * 1000)
|
163 |
+
speaker_ts.append([s, e, int(line_list[11].split("_")[-1])])
|
164 |
+
|
165 |
+
wsm = get_words_speaker_mapping(word_timestamps, speaker_ts, "start")
|
166 |
+
|
167 |
+
if language in punct_model_langs:
|
168 |
+
# restoring punctuation in the transcript to help realign the sentences
|
169 |
+
punct_model = PunctuationModel(model="kredor/punctuate-all")
|
170 |
+
words_list = list(map(lambda x: x["word"], wsm))
|
171 |
+
labled_words = punct_model.predict(words_list)
|
172 |
+
ending_puncts = ".?!"
|
173 |
+
model_puncts = ".,;:!?"
|
174 |
+
# We don't want to punctuate U.S.A. with a period. Right?
|
175 |
+
is_acronym = lambda x: re.fullmatch(r"\b(?:[a-zA-Z]\.){2,}", x)
|
176 |
+
for word_dict, labeled_tuple in zip(wsm, labled_words):
|
177 |
+
word = word_dict["word"]
|
178 |
+
if (
|
179 |
+
word
|
180 |
+
and labeled_tuple[1] in ending_puncts
|
181 |
+
and (word[-1] not in model_puncts or is_acronym(word))
|
182 |
+
):
|
183 |
+
word += labeled_tuple[1]
|
184 |
+
if word.endswith(".."):
|
185 |
+
word = word.rstrip(".")
|
186 |
+
word_dict["word"] = word
|
187 |
+
else:
|
188 |
+
logging.warning(
|
189 |
+
f"Punctuation restoration is not available for {language} language. Using the original punctuation."
|
190 |
+
)
|
191 |
+
|
192 |
+
wsm = get_realigned_ws_mapping_with_punctuation(wsm)
|
193 |
+
ssm = get_sentences_speaker_mapping(wsm, speaker_ts)
|
194 |
+
|
195 |
+
# Load the SRT file
|
196 |
+
subs = pysrt.open(args.srt)
|
197 |
+
|
198 |
+
# Base directory for the LJ Speech-like structure
|
199 |
+
base_dir = "LJ_Speech_dataset"
|
200 |
+
|
201 |
+
# Dictionary to hold audio segments and texts for easpeaker_audios_texts = {}
|
202 |
+
|
203 |
+
# Process each subtitle
|
204 |
+
for sub in subs:
|
205 |
+
start_time = (sub.start.hours * 3600 + sub.start.minutes * 60 + sub.start.seconds) * 1000 + sub.start.milliseconds
|
206 |
+
end_time = (sub.end.hours * 3600 + sub.end.minutes * 60 + sub.end.seconds) * 1000 + sub.end.milliseconds
|
207 |
+
|
208 |
+
# Extract speaker and text from the subtitle
|
209 |
+
speaker_text = sub.text.split(':')
|
210 |
+
if len(speaker_text) > 1:
|
211 |
+
speaker = speaker_text[0].strip()
|
212 |
+
text = ':'.join(speaker_text[1:]).strip()
|
213 |
+
segment = sound[start_time:end_time]
|
214 |
+
|
215 |
+
# Append or create the audio segment and text for the speaker
|
216 |
+
if speaker not in speaker_audios_texts:
|
217 |
+
speaker_audios_texts[speaker] = []
|
218 |
+
speaker_audios_texts[speaker].append((segment, text))
|
219 |
+
|
220 |
+
# Save each speaker's audio to a separate file and generate metadata
|
221 |
+
for speaker, segments_texts in speaker_audios_texts.items():
|
222 |
+
speaker_dir = os.path.join(base_dir, speaker.replace(' ', '_'))
|
223 |
+
ensure_dir(speaker_dir)
|
224 |
+
|
225 |
+
metadata_lines = []
|
226 |
+
for i, (segment, text) in enumerate(segments_texts, start=1):
|
227 |
+
filename = f"{speaker.replace(' ', '_')}_{i:03}.wav"
|
228 |
+
filepath = os.path.join(speaker_dir, filename)
|
229 |
+
segment.export(filepath, format="wav")
|
230 |
+
|
231 |
+
# Prepare metadata line (filename without extension, speaker, text)
|
232 |
+
metadata_lines.append(f"{filename[:-4]}|{speaker}|{text}")
|
233 |
+
|
234 |
+
# Save metadata to a file
|
235 |
+
metadata_file = os.path.join(speaker_dir, "metadata.csv")
|
236 |
+
with open(metadata_file, "w", encoding="utf-8") as f:
|
237 |
+
f.write("\n".join(metadata_lines))
|
238 |
+
|
239 |
+
print(f"Exported files and metadata for {speaker}")
|
240 |
+
|
241 |
+
# Move the original WAV and SRT files to the "handled" subfolder
|
242 |
+
handled_dir = "handled"
|
243 |
+
ensure_dir(handled_dir)
|
244 |
+
os.rename(args.audio, os.path.join(handled_dir, os.path.basename(args.audio)))
|
245 |
+
os.rename(args.srt, os.path.join(handled_dir, os.path.basename(args.srt)))
|
246 |
+
|
247 |
+
print(f"Moved {args.audio} and {args.srt} to the 'handled' subfolder.")
|
248 |
+
|
249 |
+
with open(f"{os.path.splitext(args.audio)[0]}.txt", "w", encoding="utf-8-sig") as f:
|
250 |
+
get_speaker_aware_transcript(ssm, f)
|
251 |
+
|
252 |
+
with open(f"{os.path.splitext(args.audio)[0]}.srt", "w", encoding="utf-8-sig") as srt:
|
253 |
+
write_srt(ssm, srt)
|
254 |
+
|
255 |
+
cleanup(temp_path)
|
bulktranscript.py
ADDED
@@ -0,0 +1,222 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
from helpers import *
|
4 |
+
from faster_whisper import WhisperModel
|
5 |
+
import whisperx
|
6 |
+
import torch
|
7 |
+
from pydub import AudioSegment
|
8 |
+
from nemo.collections.asr.models.msdd_models import NeuralDiarizer
|
9 |
+
from deepmultilingualpunctuation import PunctuationModel
|
10 |
+
import re
|
11 |
+
import logging
|
12 |
+
import shutil
|
13 |
+
|
14 |
+
mtypes = {"cpu": "int8", "cuda": "float16"}
|
15 |
+
|
16 |
+
# Initialize parser
|
17 |
+
parser = argparse.ArgumentParser()
|
18 |
+
parser.add_argument(
|
19 |
+
"-d", "--directory", help="path to the directory containing the target files", required=True
|
20 |
+
)
|
21 |
+
parser.add_argument(
|
22 |
+
"--no-stem",
|
23 |
+
action="store_false",
|
24 |
+
dest="stemming",
|
25 |
+
default=True,
|
26 |
+
help="Disables source separation."
|
27 |
+
"This helps with long files that don't contain a lot of music.",
|
28 |
+
)
|
29 |
+
parser.add_argument(
|
30 |
+
"--suppress_numerals",
|
31 |
+
action="store_true",
|
32 |
+
dest="suppress_numerals",
|
33 |
+
default=False,
|
34 |
+
help="Suppresses Numerical Digits."
|
35 |
+
"This helps the diarization accuracy but converts all digits into written text.",
|
36 |
+
)
|
37 |
+
parser.add_argument(
|
38 |
+
"--whisper-model",
|
39 |
+
dest="model_name",
|
40 |
+
default="medium.en",
|
41 |
+
help="name of the Whisper model to use",
|
42 |
+
)
|
43 |
+
parser.add_argument(
|
44 |
+
"--batch-size",
|
45 |
+
type=int,
|
46 |
+
dest="batch_size",
|
47 |
+
default=8,
|
48 |
+
help="Batch size for batched inference, reduce if you run out of memory, set to 0 for non-batched inference",
|
49 |
+
)
|
50 |
+
parser.add_argument(
|
51 |
+
"--language",
|
52 |
+
type=str,
|
53 |
+
default=None,
|
54 |
+
choices=whisper_langs,
|
55 |
+
help="Language spoken in the audio, specify None to perform language detection",
|
56 |
+
)
|
57 |
+
parser.add_argument(
|
58 |
+
"--device",
|
59 |
+
dest="device",
|
60 |
+
default="cuda" if torch.cuda.is_available() else "cpu",
|
61 |
+
help="if you have a GPU use 'cuda', otherwise 'cpu'",
|
62 |
+
)
|
63 |
+
args = parser.parse_args()
|
64 |
+
|
65 |
+
def process_file(audio_file, output_dir):
|
66 |
+
if args.stemming:
|
67 |
+
# Isolate vocals from the rest of the audio
|
68 |
+
return_code = os.system(
|
69 |
+
f'python3 -m demucs.separate -n htdemucs --two-stems=vocals "{audio_file}" -o "temp_outputs"'
|
70 |
+
)
|
71 |
+
if return_code != 0:
|
72 |
+
logging.warning(
|
73 |
+
"Source splitting failed, using original audio file. Use --no-stem argument to disable it."
|
74 |
+
)
|
75 |
+
vocal_target = audio_file
|
76 |
+
else:
|
77 |
+
vocal_target = os.path.join(
|
78 |
+
"temp_outputs",
|
79 |
+
"htdemucs",
|
80 |
+
os.path.splitext(os.path.basename(audio_file))[0],
|
81 |
+
"vocals.wav",
|
82 |
+
)
|
83 |
+
else:
|
84 |
+
vocal_target = audio_file
|
85 |
+
|
86 |
+
# Transcribe the audio file
|
87 |
+
if args.batch_size != 0:
|
88 |
+
from transcription_helpers import transcribe_batched
|
89 |
+
whisper_results, language = transcribe_batched(
|
90 |
+
vocal_target,
|
91 |
+
args.language,
|
92 |
+
args.batch_size,
|
93 |
+
args.model_name,
|
94 |
+
mtypes[args.device],
|
95 |
+
args.suppress_numerals,
|
96 |
+
args.device,
|
97 |
+
)
|
98 |
+
else:
|
99 |
+
from transcription_helpers import transcribe
|
100 |
+
whisper_results, language = transcribe(
|
101 |
+
vocal_target,
|
102 |
+
args.language,
|
103 |
+
args.model_name,
|
104 |
+
mtypes[args.device],
|
105 |
+
args.suppress_numerals,
|
106 |
+
args.device,
|
107 |
+
)
|
108 |
+
|
109 |
+
if language in wav2vec2_langs:
|
110 |
+
alignment_model, metadata = whisperx.load_align_model(
|
111 |
+
language_code=language, device=args.device
|
112 |
+
)
|
113 |
+
result_aligned = whisperx.align(
|
114 |
+
whisper_results, alignment_model, metadata, vocal_target, args.device
|
115 |
+
)
|
116 |
+
word_timestamps = filter_missing_timestamps(
|
117 |
+
result_aligned["word_segments"],
|
118 |
+
initial_timestamp=whisper_results[0].get("start"),
|
119 |
+
final_timestamp=whisper_results[-1].get("end"),
|
120 |
+
)
|
121 |
+
# clear gpu vram
|
122 |
+
del alignment_model
|
123 |
+
torch.cuda.empty_cache()
|
124 |
+
else:
|
125 |
+
assert (
|
126 |
+
args.batch_size == 0 # TODO: add a better check for word timestamps existence
|
127 |
+
), (
|
128 |
+
f"Unsupported language: {language}, use --batch_size to 0"
|
129 |
+
" to generate word timestamps using whisper directly and fix this error."
|
130 |
+
)
|
131 |
+
word_timestamps = []
|
132 |
+
for segment in whisper_results:
|
133 |
+
for word in segment["words"]:
|
134 |
+
word_timestamps.append({"word": word[2], "start": word[0], "end": word[1]})
|
135 |
+
|
136 |
+
# convert audio to mono for NeMo compatibility
|
137 |
+
sound = AudioSegment.from_file(vocal_target).set_channels(1)
|
138 |
+
temp_path = os.path.join(output_dir, "temp_outputs")
|
139 |
+
os.makedirs(temp_path, exist_ok=True)
|
140 |
+
sound.export(os.path.join(temp_path, "mono_file.wav"), format="wav")
|
141 |
+
|
142 |
+
# Initialize NeMo MSDD diarization model
|
143 |
+
msdd_model = NeuralDiarizer(cfg=create_config(temp_path)).to(args.device)
|
144 |
+
msdd_model.diarize()
|
145 |
+
del msdd_model
|
146 |
+
torch.cuda.empty_cache()
|
147 |
+
|
148 |
+
# Reading timestamps <> Speaker Labels mapping
|
149 |
+
speaker_ts = []
|
150 |
+
with open(os.path.join(temp_path, "pred_rttms", "mono_file.rttm"), "r") as f:
|
151 |
+
lines = f.readlines()
|
152 |
+
for line in lines:
|
153 |
+
line_list = line.split(" ")
|
154 |
+
s = int(float(line_list[5]) * 1000)
|
155 |
+
e = s + int(float(line_list[8]) * 1000)
|
156 |
+
speaker_ts.append([s, e, int(line_list[11].split("_")[-1])])
|
157 |
+
|
158 |
+
wsm = get_words_speaker_mapping(word_timestamps, speaker_ts, "start")
|
159 |
+
|
160 |
+
if language in punct_model_langs:
|
161 |
+
# restoring punctuation in the transcript to help realign the sentences
|
162 |
+
punct_model = PunctuationModel(model="kredor/punctuate-all")
|
163 |
+
words_list = list(map(lambda x: x["word"], wsm))
|
164 |
+
labled_words = punct_model.predict(words_list)
|
165 |
+
ending_puncts = ".?!"
|
166 |
+
model_puncts = ".,;:!?"
|
167 |
+
# We don't want to punctuate U.S.A. with a period. Right?
|
168 |
+
is_acronym = lambda x: re.fullmatch(r"\b(?:[a-zA-Z]\.){2,}", x)
|
169 |
+
for word_dict, labeled_tuple in zip(wsm, labled_words):
|
170 |
+
word = word_dict["word"]
|
171 |
+
if (
|
172 |
+
word
|
173 |
+
and labeled_tuple[1] in ending_puncts
|
174 |
+
and (word[-1] not in model_puncts or is_acronym(word))
|
175 |
+
):
|
176 |
+
word += labeled_tuple[1]
|
177 |
+
if word.endswith(".."):
|
178 |
+
word = word.rstrip(".")
|
179 |
+
word_dict["word"] = word
|
180 |
+
else:
|
181 |
+
logging.warning(
|
182 |
+
f"Punctuation restoration is not available for {language} language. Using the original punctuation."
|
183 |
+
)
|
184 |
+
|
185 |
+
wsm = get_realigned_ws_mapping_with_punctuation(wsm)
|
186 |
+
ssm = get_sentences_speaker_mapping(wsm, speaker_ts)
|
187 |
+
|
188 |
+
with open(os.path.join(output_dir, f"{os.path.splitext(os.path.basename(audio_file))[0]}.txt"), "w", encoding="utf-8-sig") as f:
|
189 |
+
get_speaker_aware_transcript(ssm, f)
|
190 |
+
|
191 |
+
with open(os.path.join(output_dir, f"{os.path.splitext(os.path.basename(audio_file))[0]}.srt"), "w", encoding="utf-8-sig") as srt:
|
192 |
+
write_srt(ssm, srt)
|
193 |
+
|
194 |
+
cleanup(temp_path)
|
195 |
+
|
196 |
+
# Set the target directory containing the .avi files
|
197 |
+
target_dir = args.directory
|
198 |
+
|
199 |
+
# Create the "done" directory in the same location as the script
|
200 |
+
script_dir = os.path.dirname(os.path.abspath(__file__))
|
201 |
+
done_dir = os.path.join(script_dir, "done")
|
202 |
+
|
203 |
+
# Iterate over the subfolders in the target directory
|
204 |
+
for root, dirs, files in os.walk(target_dir):
|
205 |
+
for file in files:
|
206 |
+
if file.endswith(".avi"):
|
207 |
+
avi_file = os.path.join(root, file)
|
208 |
+
wav_file = os.path.splitext(avi_file)[0] + ".wav"
|
209 |
+
|
210 |
+
# Extract the audio from the .avi file
|
211 |
+
os.system(f'ffmpeg -i "{avi_file}" -vn -acodec pcm_s16le -ar 16000 -ac 1 "{wav_file}"')
|
212 |
+
|
213 |
+
# Create the mirrored subfolder structure in the "done" directory
|
214 |
+
subfolder = os.path.relpath(root, target_dir)
|
215 |
+
output_dir = os.path.join(done_dir, subfolder)
|
216 |
+
os.makedirs(output_dir, exist_ok=True)
|
217 |
+
|
218 |
+
# Process the extracted .wav file
|
219 |
+
process_file(wav_file, output_dir)
|
220 |
+
|
221 |
+
# Remove the extracted .wav file
|
222 |
+
os.remove(wav_file)
|
combinesets.py
ADDED
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import shutil
|
3 |
+
|
4 |
+
# Function to ensure directory exists
|
5 |
+
def ensure_dir(directory):
|
6 |
+
if not os.path.exists(directory):
|
7 |
+
os.makedirs(directory)
|
8 |
+
|
9 |
+
# Base directory for the LJ Speech-like structure
|
10 |
+
base_dir = "LJ_Speech_dataset"
|
11 |
+
|
12 |
+
# Prompt the user to enter the name of the chosen folder
|
13 |
+
chosen_folder = input("Enter the name of the chosen folder: ")
|
14 |
+
chosen_folder_path = os.path.join(base_dir, chosen_folder)
|
15 |
+
|
16 |
+
# Check if the chosen folder exists
|
17 |
+
if not os.path.isdir(chosen_folder_path):
|
18 |
+
print("Chosen folder does not exist.")
|
19 |
+
exit(1)
|
20 |
+
|
21 |
+
# Initialize the merge folder counter
|
22 |
+
merge_folder_counter = 2
|
23 |
+
|
24 |
+
while merge_folder_counter <= 10:
|
25 |
+
# Construct the merge folder name
|
26 |
+
merge_folder = f"{chosen_folder}{merge_folder_counter}"
|
27 |
+
merge_folder_path = os.path.join(base_dir, merge_folder)
|
28 |
+
|
29 |
+
# Check if the merge folder exists
|
30 |
+
if not os.path.isdir(merge_folder_path):
|
31 |
+
# Increment the merge folder counter and continue to the next iteration
|
32 |
+
merge_folder_counter += 1
|
33 |
+
continue
|
34 |
+
|
35 |
+
# Initialize variables for renaming files
|
36 |
+
file_counter = len(os.listdir(chosen_folder_path)) // 2 + 1
|
37 |
+
metadata_lines = []
|
38 |
+
|
39 |
+
# Process the merge folder
|
40 |
+
for filename in os.listdir(merge_folder_path):
|
41 |
+
if filename.endswith(".wav"):
|
42 |
+
# Update the filename to include the merge folder name
|
43 |
+
old_filename = filename
|
44 |
+
new_filename = f"{merge_folder}_{filename}"
|
45 |
+
old_path = os.path.join(merge_folder_path, old_filename)
|
46 |
+
new_path = os.path.join(merge_folder_path, new_filename)
|
47 |
+
os.rename(old_path, new_path)
|
48 |
+
|
49 |
+
# Read the corresponding text from the metadata file
|
50 |
+
metadata_file = os.path.join(merge_folder_path, "metadata.csv")
|
51 |
+
with open(metadata_file, "r", encoding="utf-8") as f:
|
52 |
+
for line in f:
|
53 |
+
if line.startswith(old_filename[:-4]):
|
54 |
+
text = line.strip().split("|")[2]
|
55 |
+
break
|
56 |
+
|
57 |
+
# Prepare metadata line for the chosen folder
|
58 |
+
metadata_lines.append(f"{new_filename[:-4]}|{chosen_folder}|{text}")
|
59 |
+
|
60 |
+
# Copy the updated audio file to the chosen folder
|
61 |
+
shutil.copy(new_path, chosen_folder_path)
|
62 |
+
|
63 |
+
file_counter += 1
|
64 |
+
|
65 |
+
# Update the merge folder's metadata file with the new filenames
|
66 |
+
metadata_file = os.path.join(merge_folder_path, "metadata.csv")
|
67 |
+
with open(metadata_file, "r", encoding="utf-8") as f:
|
68 |
+
lines = f.readlines()
|
69 |
+
|
70 |
+
updated_lines = []
|
71 |
+
for line in lines:
|
72 |
+
parts = line.strip().split("|")
|
73 |
+
filename = parts[0]
|
74 |
+
text = parts[2]
|
75 |
+
updated_line = f"{merge_folder}_{filename}|{merge_folder}|{text}\n"
|
76 |
+
updated_lines.append(updated_line)
|
77 |
+
|
78 |
+
with open(metadata_file, "w", encoding="utf-8") as f:
|
79 |
+
f.writelines(updated_lines)
|
80 |
+
|
81 |
+
# Append the metadata lines to the chosen folder's metadata file
|
82 |
+
metadata_file = os.path.join(chosen_folder_path, "metadata.csv")
|
83 |
+
with open(metadata_file, "a", encoding="utf-8") as f:
|
84 |
+
f.write("\n".join(metadata_lines) + "\n")
|
85 |
+
|
86 |
+
# Remove the merge folder
|
87 |
+
shutil.rmtree(merge_folder_path)
|
88 |
+
|
89 |
+
print(f"Merge completed successfully for {merge_folder}.")
|
90 |
+
|
91 |
+
# Increment the merge folder counter
|
92 |
+
merge_folder_counter += 1
|
93 |
+
|
94 |
+
print("All merge operations completed.")
|
concat.py
ADDED
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
from pydub import AudioSegment
|
3 |
+
|
4 |
+
# Create an argument parser
|
5 |
+
parser = argparse.ArgumentParser(description='Concatenate two WAV files.')
|
6 |
+
parser.add_argument('--wav1', type=str, required=True, help='Path to the first WAV file')
|
7 |
+
parser.add_argument('--wav2', type=str, required=True, help='Path to the second WAV file')
|
8 |
+
args = parser.parse_args()
|
9 |
+
|
10 |
+
# Load the audio files
|
11 |
+
audio1 = AudioSegment.from_wav(args.wav1)
|
12 |
+
audio2 = AudioSegment.from_wav(args.wav2)
|
13 |
+
|
14 |
+
# Concatenate the audio files
|
15 |
+
combined_audio = audio1 + audio2
|
16 |
+
|
17 |
+
# Export the concatenated audio to a new file
|
18 |
+
output_file = 'combined_audio.wav'
|
19 |
+
combined_audio.export(output_file, format="wav")
|
20 |
+
|
21 |
+
print(f"Concatenated audio saved as {output_file}")
|
consolidate_datasets.py
ADDED
@@ -0,0 +1,110 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import csv
|
3 |
+
import shutil
|
4 |
+
from pydub import AudioSegment
|
5 |
+
import multiprocessing
|
6 |
+
|
7 |
+
def process_folder(folder_path, output_folder):
|
8 |
+
print(f"Processing folder: {folder_path}")
|
9 |
+
|
10 |
+
# Step 1: Copy wav files and metadata.csv to the output folder
|
11 |
+
metadata_file = os.path.join(folder_path, "metadata.csv")
|
12 |
+
output_metadata_file = os.path.join(output_folder, "metadata.csv")
|
13 |
+
|
14 |
+
with open(metadata_file, "r") as file, open(output_metadata_file, "w", newline="") as output_file:
|
15 |
+
reader = csv.reader(file, delimiter="|")
|
16 |
+
writer = csv.writer(output_file, delimiter="|")
|
17 |
+
|
18 |
+
for row in reader:
|
19 |
+
wav_file = os.path.join(folder_path, row[0] + ".wav")
|
20 |
+
output_wav_file = os.path.join(output_folder, row[0] + ".wav")
|
21 |
+
shutil.copy(wav_file, output_wav_file)
|
22 |
+
print(f"Copied {wav_file} to {output_wav_file}")
|
23 |
+
writer.writerow(row)
|
24 |
+
|
25 |
+
# Step 2: Rename wav files and update metadata.csv
|
26 |
+
folder_name = os.path.basename(folder_path)
|
27 |
+
temp_metadata_file = os.path.join(output_folder, "temp_metadata.csv")
|
28 |
+
|
29 |
+
with open(output_metadata_file, "r") as file, open(temp_metadata_file, "w", newline="") as temp_file:
|
30 |
+
reader = csv.reader(file, delimiter="|")
|
31 |
+
writer = csv.writer(temp_file, delimiter="|")
|
32 |
+
|
33 |
+
for row in reader:
|
34 |
+
old_wav_file = os.path.join(output_folder, row[0] + ".wav")
|
35 |
+
new_wav_file = os.path.join(output_folder, folder_name + "_" + row[0].split("_")[-1] + ".wav")
|
36 |
+
os.rename(old_wav_file, new_wav_file)
|
37 |
+
print(f"Renamed {old_wav_file} to {new_wav_file}")
|
38 |
+
|
39 |
+
row[0] = folder_name + "_" + row[0].split("_")[-1]
|
40 |
+
writer.writerow(row)
|
41 |
+
|
42 |
+
os.remove(output_metadata_file)
|
43 |
+
os.rename(temp_metadata_file, output_metadata_file)
|
44 |
+
print(f"Updated metadata.csv in {output_folder}")
|
45 |
+
|
46 |
+
def merge_folders(base_name, folder_list, output_base_folder):
|
47 |
+
merged_folder = os.path.join(output_base_folder, base_name)
|
48 |
+
os.makedirs(merged_folder, exist_ok=True)
|
49 |
+
print(f"Created merged folder: {merged_folder}")
|
50 |
+
|
51 |
+
merged_metadata_file = os.path.join(merged_folder, "metadata.csv")
|
52 |
+
with open(merged_metadata_file, "w", newline="") as merged_file:
|
53 |
+
writer = csv.writer(merged_file, delimiter="|")
|
54 |
+
|
55 |
+
for folder_path in folder_list:
|
56 |
+
metadata_file = os.path.join(output_base_folder, os.path.basename(folder_path), "metadata.csv")
|
57 |
+
with open(metadata_file, "r") as file:
|
58 |
+
reader = csv.reader(file, delimiter="|")
|
59 |
+
for row in reader:
|
60 |
+
row[1] = base_name
|
61 |
+
writer.writerow(row)
|
62 |
+
|
63 |
+
wav_files = [f for f in os.listdir(os.path.join(output_base_folder, os.path.basename(folder_path))) if f.endswith(".wav")]
|
64 |
+
for wav_file in wav_files:
|
65 |
+
old_wav_path = os.path.join(output_base_folder, os.path.basename(folder_path), wav_file)
|
66 |
+
new_wav_path = os.path.join(merged_folder, wav_file)
|
67 |
+
shutil.move(old_wav_path, new_wav_path)
|
68 |
+
print(f"Moved {old_wav_path} to {new_wav_path}")
|
69 |
+
|
70 |
+
# Remove the processed folder
|
71 |
+
shutil.rmtree(os.path.join(output_base_folder, os.path.basename(folder_path)))
|
72 |
+
|
73 |
+
print(f"Merged metadata.csv files into {merged_metadata_file}")
|
74 |
+
|
75 |
+
def process_subfolder(folder_path, output_base_folder):
|
76 |
+
output_folder = os.path.join(output_base_folder, os.path.basename(folder_path))
|
77 |
+
os.makedirs(output_folder, exist_ok=True)
|
78 |
+
process_folder(folder_path, output_folder)
|
79 |
+
|
80 |
+
if __name__ == "__main__":
|
81 |
+
# Set up input and output directories
|
82 |
+
base_folder = "/media/autometa/datapuddle/movie/whisper-diarization/LJ_Speech_dataset"
|
83 |
+
output_base_folder = "/media/autometa/datapuddle/movie/whisper-diarization/LJSpeech-dense"
|
84 |
+
|
85 |
+
# Create the output base folder if it doesn't exist
|
86 |
+
os.makedirs(output_base_folder, exist_ok=True)
|
87 |
+
|
88 |
+
# Get the list of subfolders
|
89 |
+
subfolders = [os.path.join(base_folder, folder_name) for folder_name in os.listdir(base_folder) if os.path.isdir(os.path.join(base_folder, folder_name))]
|
90 |
+
|
91 |
+
# Group subfolders by their base name and enumeration
|
92 |
+
folder_groups = {}
|
93 |
+
for subfolder in subfolders:
|
94 |
+
base_name = os.path.basename(subfolder).rstrip("0123456789")
|
95 |
+
if base_name not in folder_groups:
|
96 |
+
folder_groups[base_name] = []
|
97 |
+
folder_groups[base_name].append(subfolder)
|
98 |
+
|
99 |
+
# Process and merge each group of folders one at a time
|
100 |
+
for base_name, folder_list in folder_groups.items():
|
101 |
+
print(f"Processing group: {base_name}")
|
102 |
+
|
103 |
+
# Process each subfolder in the group
|
104 |
+
for folder_path in folder_list:
|
105 |
+
process_subfolder(folder_path, output_base_folder)
|
106 |
+
|
107 |
+
# Merge the folders in the group
|
108 |
+
merge_folders(base_name, folder_list, output_base_folder)
|
109 |
+
|
110 |
+
print("Processing complete.")
|
diarize.py
ADDED
@@ -0,0 +1,227 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import time
|
2 |
+
from datasets import Dataset
|
3 |
+
import warnings
|
4 |
+
import argparse
|
5 |
+
import os
|
6 |
+
from helpers import *
|
7 |
+
from faster_whisper import WhisperModel
|
8 |
+
import whisperx
|
9 |
+
import torch
|
10 |
+
from pydub import AudioSegment
|
11 |
+
from nemo.collections.asr.models.msdd_models import NeuralDiarizer
|
12 |
+
from deepmultilingualpunctuation import PunctuationModel
|
13 |
+
import re
|
14 |
+
import logging
|
15 |
+
|
16 |
+
mtypes = {"cpu": "int8", "cuda": "float16"}
|
17 |
+
|
18 |
+
|
19 |
+
|
20 |
+
# Initialize parser
|
21 |
+
parser = argparse.ArgumentParser()
|
22 |
+
parser.add_argument(
|
23 |
+
"-a", "--audio", help="name of the target audio file", required=True
|
24 |
+
)
|
25 |
+
parser.add_argument(
|
26 |
+
"--no-stem",
|
27 |
+
action="store_false",
|
28 |
+
dest="stemming",
|
29 |
+
default=True,
|
30 |
+
help="Disables source separation."
|
31 |
+
"This helps with long files that don't contain a lot of music.",
|
32 |
+
)
|
33 |
+
|
34 |
+
parser.add_argument(
|
35 |
+
"--suppress_numerals",
|
36 |
+
action="store_true",
|
37 |
+
dest="suppress_numerals",
|
38 |
+
default=False,
|
39 |
+
help="Suppresses Numerical Digits."
|
40 |
+
"This helps the diarization accuracy but converts all digits into written text.",
|
41 |
+
)
|
42 |
+
|
43 |
+
parser.add_argument(
|
44 |
+
"--whisper-model",
|
45 |
+
dest="model_name",
|
46 |
+
default="medium.en",
|
47 |
+
help="name of the Whisper model to use",
|
48 |
+
)
|
49 |
+
|
50 |
+
parser.add_argument(
|
51 |
+
"--batch-size",
|
52 |
+
type=int,
|
53 |
+
dest="batch_size",
|
54 |
+
default=8,
|
55 |
+
help="Batch size for batched inference, reduce if you run out of memory, set to 0 for non-batched inference",
|
56 |
+
)
|
57 |
+
|
58 |
+
parser.add_argument(
|
59 |
+
"--language",
|
60 |
+
type=str,
|
61 |
+
default=None,
|
62 |
+
choices=whisper_langs,
|
63 |
+
help="Language spoken in the audio, specify None to perform language detection",
|
64 |
+
)
|
65 |
+
|
66 |
+
parser.add_argument(
|
67 |
+
"--device",
|
68 |
+
dest="device",
|
69 |
+
default="cuda" if torch.cuda.is_available() else "cpu",
|
70 |
+
help="if you have a GPU use 'cuda', otherwise 'cpu'",
|
71 |
+
)
|
72 |
+
|
73 |
+
args = parser.parse_args()
|
74 |
+
|
75 |
+
if args.stemming:
|
76 |
+
# Isolate vocals from the rest of the audio
|
77 |
+
|
78 |
+
return_code = os.system(
|
79 |
+
f'python3 -m demucs.separate -n htdemucs --two-stems=vocals "{args.audio}" -o "temp_outputs"'
|
80 |
+
)
|
81 |
+
|
82 |
+
if return_code != 0:
|
83 |
+
logging.warning(
|
84 |
+
"Source splitting failed, using original audio file. Use --no-stem argument to disable it."
|
85 |
+
)
|
86 |
+
vocal_target = args.audio
|
87 |
+
else:
|
88 |
+
vocal_target = os.path.join(
|
89 |
+
"temp_outputs",
|
90 |
+
"htdemucs",
|
91 |
+
os.path.splitext(os.path.basename(args.audio))[0],
|
92 |
+
"vocals.wav",
|
93 |
+
)
|
94 |
+
else:
|
95 |
+
vocal_target = args.audio
|
96 |
+
|
97 |
+
|
98 |
+
# Transcribe the audio file
|
99 |
+
if args.batch_size != 0:
|
100 |
+
from transcription_helpers import transcribe_batched
|
101 |
+
|
102 |
+
whisper_results, language = transcribe_batched(
|
103 |
+
vocal_target,
|
104 |
+
args.language,
|
105 |
+
args.batch_size,
|
106 |
+
args.model_name,
|
107 |
+
mtypes[args.device],
|
108 |
+
args.suppress_numerals,
|
109 |
+
args.device,
|
110 |
+
)
|
111 |
+
else:
|
112 |
+
from transcription_helpers import transcribe
|
113 |
+
|
114 |
+
whisper_results, language = transcribe(
|
115 |
+
vocal_target,
|
116 |
+
args.language,
|
117 |
+
args.model_name,
|
118 |
+
mtypes[args.device],
|
119 |
+
args.suppress_numerals,
|
120 |
+
args.device,
|
121 |
+
)
|
122 |
+
|
123 |
+
if language in wav2vec2_langs:
|
124 |
+
alignment_model, metadata = whisperx.load_align_model(
|
125 |
+
language_code=language, device=args.device
|
126 |
+
)
|
127 |
+
result_aligned = whisperx.align(
|
128 |
+
whisper_results, alignment_model, metadata, vocal_target, args.device
|
129 |
+
)
|
130 |
+
word_timestamps = filter_missing_timestamps(
|
131 |
+
result_aligned["word_segments"],
|
132 |
+
initial_timestamp=whisper_results[0].get("start"),
|
133 |
+
final_timestamp=whisper_results[-1].get("end"),
|
134 |
+
)
|
135 |
+
# clear gpu vram
|
136 |
+
del alignment_model
|
137 |
+
torch.cuda.empty_cache()
|
138 |
+
else:
|
139 |
+
assert (
|
140 |
+
args.batch_size == 0 # TODO: add a better check for word timestamps existence
|
141 |
+
), (
|
142 |
+
f"Unsupported language: {language}, use --batch_size to 0"
|
143 |
+
" to generate word timestamps using whisper directly and fix this error."
|
144 |
+
)
|
145 |
+
word_timestamps = []
|
146 |
+
for segment in whisper_results:
|
147 |
+
for word in segment["words"]:
|
148 |
+
word_timestamps.append({"word": word[2], "start": word[0], "end": word[1]})
|
149 |
+
|
150 |
+
|
151 |
+
# convert audio to mono for NeMo combatibility
|
152 |
+
sound = AudioSegment.from_file(vocal_target).set_channels(1)
|
153 |
+
ROOT = os.getcwd()
|
154 |
+
temp_path = os.path.join(ROOT, "temp_outputs")
|
155 |
+
os.makedirs(temp_path, exist_ok=True)
|
156 |
+
sound.export(os.path.join(temp_path, "mono_file.wav"), format="wav")
|
157 |
+
|
158 |
+
# Initialize NeMo MSDD diarization model
|
159 |
+
msdd_model = NeuralDiarizer(cfg=create_config(temp_path)).to(args.device)
|
160 |
+
msdd_model.diarize()
|
161 |
+
|
162 |
+
del msdd_model
|
163 |
+
torch.cuda.empty_cache()
|
164 |
+
|
165 |
+
# Reading timestamps <> Speaker Labels mapping
|
166 |
+
|
167 |
+
|
168 |
+
speaker_ts = []
|
169 |
+
with open(os.path.join(temp_path, "pred_rttms", "mono_file.rttm"), "r") as f:
|
170 |
+
lines = f.readlines()
|
171 |
+
for line in lines:
|
172 |
+
line_list = line.split(" ")
|
173 |
+
s = int(float(line_list[5]) * 1000)
|
174 |
+
e = s + int(float(line_list[8]) * 1000)
|
175 |
+
speaker_ts.append([s, e, int(line_list[11].split("_")[-1])])
|
176 |
+
|
177 |
+
wsm = get_words_speaker_mapping(word_timestamps, speaker_ts, "start")
|
178 |
+
|
179 |
+
if language in punct_model_langs:
|
180 |
+
# restoring punctuation in the transcript to help realign the sentences
|
181 |
+
punct_model = PunctuationModel(model="kredor/punctuate-all")
|
182 |
+
words_list = list(map(lambda x: x["word"], wsm))
|
183 |
+
|
184 |
+
# Use the pipe method directly on the words_list
|
185 |
+
while True:
|
186 |
+
try:
|
187 |
+
labled_words = punct_model.pipe(words_list)
|
188 |
+
break
|
189 |
+
except ValueError as e:
|
190 |
+
if str(e) == "Queue is full! Please try again.":
|
191 |
+
print("Queue is full. Retrying in 1 second...")
|
192 |
+
time.sleep(1)
|
193 |
+
else:
|
194 |
+
raise e
|
195 |
+
|
196 |
+
ending_puncts = ".?!"
|
197 |
+
model_puncts = ".,;:!?"
|
198 |
+
# We don't want to punctuate U.S.A. with a period. Right?
|
199 |
+
is_acronym = lambda x: re.fullmatch(r"\b(?:[a-zA-Z]\.){2,}", x)
|
200 |
+
for i, labeled_tuple in enumerate(labled_words):
|
201 |
+
word = wsm[i]["word"]
|
202 |
+
if (
|
203 |
+
word
|
204 |
+
and labeled_tuple
|
205 |
+
and "entity" in labeled_tuple[0]
|
206 |
+
and labeled_tuple[0]["entity"] in ending_puncts
|
207 |
+
and (word[-1] not in model_puncts or is_acronym(word))
|
208 |
+
):
|
209 |
+
word += labeled_tuple[0]["entity"]
|
210 |
+
if word.endswith(".."):
|
211 |
+
word = word.rstrip(".")
|
212 |
+
wsm[i]["word"] = word
|
213 |
+
else:
|
214 |
+
logging.warning(
|
215 |
+
f"Punctuation restoration is not available for {language} language. Using the original punctuation."
|
216 |
+
)
|
217 |
+
|
218 |
+
wsm = get_realigned_ws_mapping_with_punctuation(wsm)
|
219 |
+
ssm = get_sentences_speaker_mapping(wsm, speaker_ts)
|
220 |
+
|
221 |
+
with open(f"{os.path.splitext(args.audio)[0]}.txt", "w", encoding="utf-8-sig") as f:
|
222 |
+
get_speaker_aware_transcript(ssm, f)
|
223 |
+
|
224 |
+
with open(f"{os.path.splitext(args.audio)[0]}.srt", "w", encoding="utf-8-sig") as srt:
|
225 |
+
write_srt(ssm, srt)
|
226 |
+
|
227 |
+
cleanup(temp_path)
|
diarize_parallel.py
ADDED
@@ -0,0 +1,203 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
from helpers import *
|
4 |
+
from faster_whisper import WhisperModel
|
5 |
+
import whisperx
|
6 |
+
import torch
|
7 |
+
from deepmultilingualpunctuation import PunctuationModel
|
8 |
+
import re
|
9 |
+
import subprocess
|
10 |
+
import logging
|
11 |
+
|
12 |
+
mtypes = {"cpu": "int8", "cuda": "float16"}
|
13 |
+
|
14 |
+
# Initialize parser
|
15 |
+
parser = argparse.ArgumentParser()
|
16 |
+
parser.add_argument(
|
17 |
+
"-a", "--audio", help="name of the target audio file", required=True
|
18 |
+
)
|
19 |
+
parser.add_argument(
|
20 |
+
"--no-stem",
|
21 |
+
action="store_false",
|
22 |
+
dest="stemming",
|
23 |
+
default=True,
|
24 |
+
help="Disables source separation."
|
25 |
+
"This helps with long files that don't contain a lot of music.",
|
26 |
+
)
|
27 |
+
|
28 |
+
parser.add_argument(
|
29 |
+
"--suppress_numerals",
|
30 |
+
action="store_true",
|
31 |
+
dest="suppress_numerals",
|
32 |
+
default=False,
|
33 |
+
help="Suppresses Numerical Digits."
|
34 |
+
"This helps the diarization accuracy but converts all digits into written text.",
|
35 |
+
)
|
36 |
+
|
37 |
+
parser.add_argument(
|
38 |
+
"--whisper-model",
|
39 |
+
dest="model_name",
|
40 |
+
default="medium.en",
|
41 |
+
help="name of the Whisper model to use",
|
42 |
+
)
|
43 |
+
|
44 |
+
parser.add_argument(
|
45 |
+
"--batch-size",
|
46 |
+
type=int,
|
47 |
+
dest="batch_size",
|
48 |
+
default=8,
|
49 |
+
help="Batch size for batched inference, reduce if you run out of memory, set to 0 for non-batched inference",
|
50 |
+
)
|
51 |
+
|
52 |
+
parser.add_argument(
|
53 |
+
"--language",
|
54 |
+
type=str,
|
55 |
+
default=None,
|
56 |
+
choices=whisper_langs,
|
57 |
+
help="Language spoken in the audio, specify None to perform language detection",
|
58 |
+
)
|
59 |
+
|
60 |
+
parser.add_argument(
|
61 |
+
"--device",
|
62 |
+
dest="device",
|
63 |
+
default="cuda" if torch.cuda.is_available() else "cpu",
|
64 |
+
help="if you have a GPU use 'cuda', otherwise 'cpu'",
|
65 |
+
)
|
66 |
+
|
67 |
+
args = parser.parse_args()
|
68 |
+
|
69 |
+
if args.stemming:
|
70 |
+
# Isolate vocals from the rest of the audio
|
71 |
+
|
72 |
+
return_code = os.system(
|
73 |
+
f'python3 -m demucs.separate -n htdemucs --two-stems=vocals "{args.audio}" -o "temp_outputs"'
|
74 |
+
)
|
75 |
+
|
76 |
+
if return_code != 0:
|
77 |
+
logging.warning(
|
78 |
+
"Source splitting failed, using original audio file. Use --no-stem argument to disable it."
|
79 |
+
)
|
80 |
+
vocal_target = args.audio
|
81 |
+
else:
|
82 |
+
vocal_target = os.path.join(
|
83 |
+
"temp_outputs",
|
84 |
+
"htdemucs",
|
85 |
+
os.path.splitext(os.path.basename(args.audio))[0],
|
86 |
+
"vocals.wav",
|
87 |
+
)
|
88 |
+
else:
|
89 |
+
vocal_target = args.audio
|
90 |
+
|
91 |
+
logging.info("Starting Nemo process with vocal_target: ", vocal_target)
|
92 |
+
nemo_process = subprocess.Popen(
|
93 |
+
["python3", "nemo_process.py", "-a", vocal_target, "--device", args.device],
|
94 |
+
)
|
95 |
+
# Transcribe the audio file
|
96 |
+
if args.batch_size != 0:
|
97 |
+
from transcription_helpers import transcribe_batched
|
98 |
+
|
99 |
+
whisper_results, language = transcribe_batched(
|
100 |
+
vocal_target,
|
101 |
+
args.language,
|
102 |
+
args.batch_size,
|
103 |
+
args.model_name,
|
104 |
+
mtypes[args.device],
|
105 |
+
args.suppress_numerals,
|
106 |
+
args.device,
|
107 |
+
)
|
108 |
+
else:
|
109 |
+
from transcription_helpers import transcribe
|
110 |
+
|
111 |
+
whisper_results, language = transcribe(
|
112 |
+
vocal_target,
|
113 |
+
args.language,
|
114 |
+
args.model_name,
|
115 |
+
mtypes[args.device],
|
116 |
+
args.suppress_numerals,
|
117 |
+
args.device,
|
118 |
+
)
|
119 |
+
|
120 |
+
if language in wav2vec2_langs:
|
121 |
+
alignment_model, metadata = whisperx.load_align_model(
|
122 |
+
language_code=language, device=args.device
|
123 |
+
)
|
124 |
+
result_aligned = whisperx.align(
|
125 |
+
whisper_results, alignment_model, metadata, vocal_target, args.device
|
126 |
+
)
|
127 |
+
word_timestamps = filter_missing_timestamps(
|
128 |
+
result_aligned["word_segments"],
|
129 |
+
initial_timestamp=whisper_results[0].get("start"),
|
130 |
+
final_timestamp=whisper_results[-1].get("end"),
|
131 |
+
)
|
132 |
+
# clear gpu vram
|
133 |
+
del alignment_model
|
134 |
+
torch.cuda.empty_cache()
|
135 |
+
else:
|
136 |
+
assert (
|
137 |
+
args.batch_size == 0 # TODO: add a better check for word timestamps existence
|
138 |
+
), (
|
139 |
+
f"Unsupported language: {language}, use --batch_size to 0"
|
140 |
+
" to generate word timestamps using whisper directly and fix this error."
|
141 |
+
)
|
142 |
+
word_timestamps = []
|
143 |
+
for segment in whisper_results:
|
144 |
+
for word in segment["words"]:
|
145 |
+
word_timestamps.append({"word": word[2], "start": word[0], "end": word[1]})
|
146 |
+
|
147 |
+
# Reading timestamps <> Speaker Labels mapping
|
148 |
+
nemo_process.communicate()
|
149 |
+
ROOT = os.getcwd()
|
150 |
+
temp_path = os.path.join(ROOT, "temp_outputs")
|
151 |
+
|
152 |
+
speaker_ts = []
|
153 |
+
with open(os.path.join(temp_path, "pred_rttms", "mono_file.rttm"), "r") as f:
|
154 |
+
lines = f.readlines()
|
155 |
+
for line in lines:
|
156 |
+
line_list = line.split(" ")
|
157 |
+
s = int(float(line_list[5]) * 1000)
|
158 |
+
e = s + int(float(line_list[8]) * 1000)
|
159 |
+
speaker_ts.append([s, e, int(line_list[11].split("_")[-1])])
|
160 |
+
|
161 |
+
wsm = get_words_speaker_mapping(word_timestamps, speaker_ts, "start")
|
162 |
+
|
163 |
+
if language in punct_model_langs:
|
164 |
+
# restoring punctuation in the transcript to help realign the sentences
|
165 |
+
punct_model = PunctuationModel(model="kredor/punctuate-all")
|
166 |
+
|
167 |
+
words_list = list(map(lambda x: x["word"], wsm))
|
168 |
+
|
169 |
+
labled_words = punct_model.predict(words_list)
|
170 |
+
|
171 |
+
ending_puncts = ".?!"
|
172 |
+
model_puncts = ".,;:!?"
|
173 |
+
|
174 |
+
# We don't want to punctuate U.S.A. with a period. Right?
|
175 |
+
is_acronym = lambda x: re.fullmatch(r"\b(?:[a-zA-Z]\.){2,}", x)
|
176 |
+
|
177 |
+
for word_dict, labeled_tuple in zip(wsm, labled_words):
|
178 |
+
word = word_dict["word"]
|
179 |
+
if (
|
180 |
+
word
|
181 |
+
and labeled_tuple[1] in ending_puncts
|
182 |
+
and (word[-1] not in model_puncts or is_acronym(word))
|
183 |
+
):
|
184 |
+
word += labeled_tuple[1]
|
185 |
+
if word.endswith(".."):
|
186 |
+
word = word.rstrip(".")
|
187 |
+
word_dict["word"] = word
|
188 |
+
|
189 |
+
else:
|
190 |
+
logging.warning(
|
191 |
+
f"Punctuation restoration is not available for {language} language. Using the original punctuation."
|
192 |
+
)
|
193 |
+
|
194 |
+
wsm = get_realigned_ws_mapping_with_punctuation(wsm)
|
195 |
+
ssm = get_sentences_speaker_mapping(wsm, speaker_ts)
|
196 |
+
|
197 |
+
with open(f"{os.path.splitext(args.audio)[0]}.txt", "w", encoding="utf-8-sig") as f:
|
198 |
+
get_speaker_aware_transcript(ssm, f)
|
199 |
+
|
200 |
+
with open(f"{os.path.splitext(args.audio)[0]}.srt", "w", encoding="utf-8-sig") as srt:
|
201 |
+
write_srt(ssm, srt)
|
202 |
+
|
203 |
+
cleanup(temp_path)
|
helpers.py
ADDED
@@ -0,0 +1,392 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import wget
|
3 |
+
from omegaconf import OmegaConf
|
4 |
+
import json
|
5 |
+
import shutil
|
6 |
+
import nltk
|
7 |
+
from whisperx.alignment import DEFAULT_ALIGN_MODELS_HF, DEFAULT_ALIGN_MODELS_TORCH
|
8 |
+
import logging
|
9 |
+
from whisperx.utils import LANGUAGES, TO_LANGUAGE_CODE
|
10 |
+
|
11 |
+
punct_model_langs = [
|
12 |
+
"en",
|
13 |
+
"fr",
|
14 |
+
"de",
|
15 |
+
"es",
|
16 |
+
"it",
|
17 |
+
"nl",
|
18 |
+
"pt",
|
19 |
+
"bg",
|
20 |
+
"pl",
|
21 |
+
"cs",
|
22 |
+
"sk",
|
23 |
+
"sl",
|
24 |
+
]
|
25 |
+
wav2vec2_langs = list(DEFAULT_ALIGN_MODELS_TORCH.keys()) + list(
|
26 |
+
DEFAULT_ALIGN_MODELS_HF.keys()
|
27 |
+
)
|
28 |
+
|
29 |
+
whisper_langs = sorted(LANGUAGES.keys()) + sorted(
|
30 |
+
[k.title() for k in TO_LANGUAGE_CODE.keys()]
|
31 |
+
)
|
32 |
+
|
33 |
+
|
34 |
+
def create_config(output_dir):
|
35 |
+
DOMAIN_TYPE = "telephonic" # Can be meeting, telephonic, or general based on domain type of the audio file
|
36 |
+
CONFIG_LOCAL_DIRECTORY = "nemo_msdd_configs"
|
37 |
+
CONFIG_FILE_NAME = f"diar_infer_{DOMAIN_TYPE}.yaml"
|
38 |
+
MODEL_CONFIG_PATH = os.path.join(CONFIG_LOCAL_DIRECTORY, CONFIG_FILE_NAME)
|
39 |
+
if not os.path.exists(MODEL_CONFIG_PATH):
|
40 |
+
os.makedirs(CONFIG_LOCAL_DIRECTORY, exist_ok=True)
|
41 |
+
CONFIG_URL = f"https://raw.githubusercontent.com/NVIDIA/NeMo/main/examples/speaker_tasks/diarization/conf/inference/{CONFIG_FILE_NAME}"
|
42 |
+
MODEL_CONFIG_PATH = wget.download(CONFIG_URL, MODEL_CONFIG_PATH)
|
43 |
+
|
44 |
+
config = OmegaConf.load(MODEL_CONFIG_PATH)
|
45 |
+
|
46 |
+
data_dir = os.path.join(output_dir, "data")
|
47 |
+
os.makedirs(data_dir, exist_ok=True)
|
48 |
+
|
49 |
+
meta = {
|
50 |
+
"audio_filepath": os.path.join(output_dir, "mono_file.wav"),
|
51 |
+
"offset": 0,
|
52 |
+
"duration": None,
|
53 |
+
"label": "infer",
|
54 |
+
"text": "-",
|
55 |
+
"rttm_filepath": None,
|
56 |
+
"uem_filepath": None,
|
57 |
+
}
|
58 |
+
with open(os.path.join(data_dir, "input_manifest.json"), "w") as fp:
|
59 |
+
json.dump(meta, fp)
|
60 |
+
fp.write("\n")
|
61 |
+
|
62 |
+
pretrained_vad = "vad_multilingual_marblenet"
|
63 |
+
pretrained_speaker_model = "titanet_large"
|
64 |
+
config.num_workers = 0
|
65 |
+
config.diarizer.manifest_filepath = os.path.join(data_dir, "input_manifest.json")
|
66 |
+
config.diarizer.out_dir = (
|
67 |
+
output_dir # Directory to store intermediate files and prediction outputs
|
68 |
+
)
|
69 |
+
|
70 |
+
config.diarizer.speaker_embeddings.model_path = pretrained_speaker_model
|
71 |
+
config.diarizer.oracle_vad = (
|
72 |
+
False # compute VAD provided with model_path to vad config
|
73 |
+
)
|
74 |
+
config.diarizer.clustering.parameters.oracle_num_speakers = False
|
75 |
+
|
76 |
+
# Here, we use our in-house pretrained NeMo VAD model
|
77 |
+
config.diarizer.vad.model_path = pretrained_vad
|
78 |
+
config.diarizer.vad.parameters.onset = 0.8
|
79 |
+
config.diarizer.vad.parameters.offset = 0.6
|
80 |
+
config.diarizer.vad.parameters.pad_offset = -0.05
|
81 |
+
config.diarizer.msdd_model.model_path = (
|
82 |
+
"diar_msdd_telephonic" # Telephonic speaker diarization model
|
83 |
+
)
|
84 |
+
|
85 |
+
return config
|
86 |
+
|
87 |
+
|
88 |
+
def get_word_ts_anchor(s, e, option="start"):
|
89 |
+
if option == "end":
|
90 |
+
return e
|
91 |
+
elif option == "mid":
|
92 |
+
return (s + e) / 2
|
93 |
+
return s
|
94 |
+
|
95 |
+
|
96 |
+
def get_words_speaker_mapping(wrd_ts, spk_ts, word_anchor_option="start"):
|
97 |
+
s, e, sp = spk_ts[0]
|
98 |
+
wrd_pos, turn_idx = 0, 0
|
99 |
+
wrd_spk_mapping = []
|
100 |
+
for wrd_dict in wrd_ts:
|
101 |
+
ws, we, wrd = (
|
102 |
+
int(wrd_dict["start"] * 1000),
|
103 |
+
int(wrd_dict["end"] * 1000),
|
104 |
+
wrd_dict["word"],
|
105 |
+
)
|
106 |
+
wrd_pos = get_word_ts_anchor(ws, we, word_anchor_option)
|
107 |
+
while wrd_pos > float(e):
|
108 |
+
turn_idx += 1
|
109 |
+
turn_idx = min(turn_idx, len(spk_ts) - 1)
|
110 |
+
s, e, sp = spk_ts[turn_idx]
|
111 |
+
if turn_idx == len(spk_ts) - 1:
|
112 |
+
e = get_word_ts_anchor(ws, we, option="end")
|
113 |
+
wrd_spk_mapping.append(
|
114 |
+
{"word": wrd, "start_time": ws, "end_time": we, "speaker": sp}
|
115 |
+
)
|
116 |
+
return wrd_spk_mapping
|
117 |
+
|
118 |
+
|
119 |
+
sentence_ending_punctuations = ".?!"
|
120 |
+
|
121 |
+
|
122 |
+
def get_first_word_idx_of_sentence(word_idx, word_list, speaker_list, max_words):
|
123 |
+
is_word_sentence_end = (
|
124 |
+
lambda x: x >= 0 and word_list[x][-1] in sentence_ending_punctuations
|
125 |
+
)
|
126 |
+
left_idx = word_idx
|
127 |
+
while (
|
128 |
+
left_idx > 0
|
129 |
+
and word_idx - left_idx < max_words
|
130 |
+
and speaker_list[left_idx - 1] == speaker_list[left_idx]
|
131 |
+
and not is_word_sentence_end(left_idx - 1)
|
132 |
+
):
|
133 |
+
left_idx -= 1
|
134 |
+
|
135 |
+
return left_idx if left_idx == 0 or is_word_sentence_end(left_idx - 1) else -1
|
136 |
+
|
137 |
+
|
138 |
+
def get_last_word_idx_of_sentence(word_idx, word_list, max_words):
|
139 |
+
is_word_sentence_end = (
|
140 |
+
lambda x: x >= 0 and word_list[x][-1] in sentence_ending_punctuations
|
141 |
+
)
|
142 |
+
right_idx = word_idx
|
143 |
+
while (
|
144 |
+
right_idx < len(word_list)
|
145 |
+
and right_idx - word_idx < max_words
|
146 |
+
and not is_word_sentence_end(right_idx)
|
147 |
+
):
|
148 |
+
right_idx += 1
|
149 |
+
|
150 |
+
return (
|
151 |
+
right_idx
|
152 |
+
if right_idx == len(word_list) - 1 or is_word_sentence_end(right_idx)
|
153 |
+
else -1
|
154 |
+
)
|
155 |
+
|
156 |
+
|
157 |
+
def get_realigned_ws_mapping_with_punctuation(
|
158 |
+
word_speaker_mapping, max_words_in_sentence=50
|
159 |
+
):
|
160 |
+
is_word_sentence_end = (
|
161 |
+
lambda x: x >= 0
|
162 |
+
and word_speaker_mapping[x]["word"][-1] in sentence_ending_punctuations
|
163 |
+
)
|
164 |
+
wsp_len = len(word_speaker_mapping)
|
165 |
+
|
166 |
+
words_list, speaker_list = [], []
|
167 |
+
for k, line_dict in enumerate(word_speaker_mapping):
|
168 |
+
word, speaker = line_dict["word"], line_dict["speaker"]
|
169 |
+
words_list.append(word)
|
170 |
+
speaker_list.append(speaker)
|
171 |
+
|
172 |
+
k = 0
|
173 |
+
while k < len(word_speaker_mapping):
|
174 |
+
line_dict = word_speaker_mapping[k]
|
175 |
+
if (
|
176 |
+
k < wsp_len - 1
|
177 |
+
and speaker_list[k] != speaker_list[k + 1]
|
178 |
+
and not is_word_sentence_end(k)
|
179 |
+
):
|
180 |
+
left_idx = get_first_word_idx_of_sentence(
|
181 |
+
k, words_list, speaker_list, max_words_in_sentence
|
182 |
+
)
|
183 |
+
right_idx = (
|
184 |
+
get_last_word_idx_of_sentence(
|
185 |
+
k, words_list, max_words_in_sentence - k + left_idx - 1
|
186 |
+
)
|
187 |
+
if left_idx > -1
|
188 |
+
else -1
|
189 |
+
)
|
190 |
+
if min(left_idx, right_idx) == -1:
|
191 |
+
k += 1
|
192 |
+
continue
|
193 |
+
|
194 |
+
spk_labels = speaker_list[left_idx : right_idx + 1]
|
195 |
+
mod_speaker = max(set(spk_labels), key=spk_labels.count)
|
196 |
+
if spk_labels.count(mod_speaker) < len(spk_labels) // 2:
|
197 |
+
k += 1
|
198 |
+
continue
|
199 |
+
|
200 |
+
speaker_list[left_idx : right_idx + 1] = [mod_speaker] * (
|
201 |
+
right_idx - left_idx + 1
|
202 |
+
)
|
203 |
+
k = right_idx
|
204 |
+
|
205 |
+
k += 1
|
206 |
+
|
207 |
+
k, realigned_list = 0, []
|
208 |
+
while k < len(word_speaker_mapping):
|
209 |
+
line_dict = word_speaker_mapping[k].copy()
|
210 |
+
line_dict["speaker"] = speaker_list[k]
|
211 |
+
realigned_list.append(line_dict)
|
212 |
+
k += 1
|
213 |
+
|
214 |
+
return realigned_list
|
215 |
+
|
216 |
+
|
217 |
+
def get_sentences_speaker_mapping(word_speaker_mapping, spk_ts):
|
218 |
+
sentence_checker = nltk.tokenize.PunktSentenceTokenizer().text_contains_sentbreak
|
219 |
+
s, e, spk = spk_ts[0]
|
220 |
+
prev_spk = spk
|
221 |
+
|
222 |
+
snts = []
|
223 |
+
snt = {"speaker": f"Speaker {spk}", "start_time": s, "end_time": e, "text": ""}
|
224 |
+
|
225 |
+
for wrd_dict in word_speaker_mapping:
|
226 |
+
wrd, spk = wrd_dict["word"], wrd_dict["speaker"]
|
227 |
+
s, e = wrd_dict["start_time"], wrd_dict["end_time"]
|
228 |
+
if spk != prev_spk or sentence_checker(snt["text"] + " " + wrd):
|
229 |
+
snts.append(snt)
|
230 |
+
snt = {
|
231 |
+
"speaker": f"Speaker {spk}",
|
232 |
+
"start_time": s,
|
233 |
+
"end_time": e,
|
234 |
+
"text": "",
|
235 |
+
}
|
236 |
+
else:
|
237 |
+
snt["end_time"] = e
|
238 |
+
snt["text"] += wrd + " "
|
239 |
+
prev_spk = spk
|
240 |
+
|
241 |
+
snts.append(snt)
|
242 |
+
return snts
|
243 |
+
|
244 |
+
|
245 |
+
def get_speaker_aware_transcript(sentences_speaker_mapping, f):
|
246 |
+
previous_speaker = sentences_speaker_mapping[0]["speaker"]
|
247 |
+
f.write(f"{previous_speaker}: ")
|
248 |
+
|
249 |
+
for sentence_dict in sentences_speaker_mapping:
|
250 |
+
speaker = sentence_dict["speaker"]
|
251 |
+
sentence = sentence_dict["text"]
|
252 |
+
|
253 |
+
# If this speaker doesn't match the previous one, start a new paragraph
|
254 |
+
if speaker != previous_speaker:
|
255 |
+
f.write(f"\n\n{speaker}: ")
|
256 |
+
previous_speaker = speaker
|
257 |
+
|
258 |
+
# No matter what, write the current sentence
|
259 |
+
f.write(sentence + " ")
|
260 |
+
|
261 |
+
|
262 |
+
def format_timestamp(
|
263 |
+
milliseconds: float, always_include_hours: bool = False, decimal_marker: str = "."
|
264 |
+
):
|
265 |
+
assert milliseconds >= 0, "non-negative timestamp expected"
|
266 |
+
|
267 |
+
hours = milliseconds // 3_600_000
|
268 |
+
milliseconds -= hours * 3_600_000
|
269 |
+
|
270 |
+
minutes = milliseconds // 60_000
|
271 |
+
milliseconds -= minutes * 60_000
|
272 |
+
|
273 |
+
seconds = milliseconds // 1_000
|
274 |
+
milliseconds -= seconds * 1_000
|
275 |
+
|
276 |
+
hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
|
277 |
+
return (
|
278 |
+
f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
|
279 |
+
)
|
280 |
+
|
281 |
+
|
282 |
+
def write_srt(transcript, file):
|
283 |
+
"""
|
284 |
+
Write a transcript to a file in SRT format.
|
285 |
+
|
286 |
+
"""
|
287 |
+
for i, segment in enumerate(transcript, start=1):
|
288 |
+
# write srt lines
|
289 |
+
print(
|
290 |
+
f"{i}\n"
|
291 |
+
f"{format_timestamp(segment['start_time'], always_include_hours=True, decimal_marker=',')} --> "
|
292 |
+
f"{format_timestamp(segment['end_time'], always_include_hours=True, decimal_marker=',')}\n"
|
293 |
+
f"{segment['speaker']}: {segment['text'].strip().replace('-->', '->')}\n",
|
294 |
+
file=file,
|
295 |
+
flush=True,
|
296 |
+
)
|
297 |
+
|
298 |
+
|
299 |
+
def find_numeral_symbol_tokens(tokenizer):
|
300 |
+
numeral_symbol_tokens = [
|
301 |
+
-1,
|
302 |
+
]
|
303 |
+
for token, token_id in tokenizer.get_vocab().items():
|
304 |
+
has_numeral_symbol = any(c in "0123456789%$£" for c in token)
|
305 |
+
if has_numeral_symbol:
|
306 |
+
numeral_symbol_tokens.append(token_id)
|
307 |
+
return numeral_symbol_tokens
|
308 |
+
|
309 |
+
|
310 |
+
def _get_next_start_timestamp(word_timestamps, current_word_index, final_timestamp):
|
311 |
+
# if current word is the last word
|
312 |
+
if current_word_index == len(word_timestamps) - 1:
|
313 |
+
return word_timestamps[current_word_index]["start"]
|
314 |
+
|
315 |
+
next_word_index = current_word_index + 1
|
316 |
+
while current_word_index < len(word_timestamps) - 1:
|
317 |
+
if word_timestamps[next_word_index].get("start") is None:
|
318 |
+
# if next word doesn't have a start timestamp
|
319 |
+
# merge it with the current word and delete it
|
320 |
+
word_timestamps[current_word_index]["word"] += (
|
321 |
+
" " + word_timestamps[next_word_index]["word"]
|
322 |
+
)
|
323 |
+
|
324 |
+
word_timestamps[next_word_index]["word"] = None
|
325 |
+
next_word_index += 1
|
326 |
+
if next_word_index == len(word_timestamps):
|
327 |
+
return final_timestamp
|
328 |
+
|
329 |
+
else:
|
330 |
+
return word_timestamps[next_word_index]["start"]
|
331 |
+
|
332 |
+
|
333 |
+
def filter_missing_timestamps(
|
334 |
+
word_timestamps, initial_timestamp=0, final_timestamp=None
|
335 |
+
):
|
336 |
+
# handle the first and last word
|
337 |
+
if word_timestamps[0].get("start") is None:
|
338 |
+
word_timestamps[0]["start"] = (
|
339 |
+
initial_timestamp if initial_timestamp is not None else 0
|
340 |
+
)
|
341 |
+
word_timestamps[0]["end"] = _get_next_start_timestamp(
|
342 |
+
word_timestamps, 0, final_timestamp
|
343 |
+
)
|
344 |
+
|
345 |
+
result = [
|
346 |
+
word_timestamps[0],
|
347 |
+
]
|
348 |
+
|
349 |
+
for i, ws in enumerate(word_timestamps[1:], start=1):
|
350 |
+
# if ws doesn't have a start and end
|
351 |
+
# use the previous end as start and next start as end
|
352 |
+
if ws.get("start") is None and ws.get("word") is not None:
|
353 |
+
ws["start"] = word_timestamps[i - 1]["end"]
|
354 |
+
ws["end"] = _get_next_start_timestamp(word_timestamps, i, final_timestamp)
|
355 |
+
|
356 |
+
if ws["word"] is not None:
|
357 |
+
result.append(ws)
|
358 |
+
return result
|
359 |
+
|
360 |
+
|
361 |
+
def cleanup(path: str):
|
362 |
+
"""path could either be relative or absolute."""
|
363 |
+
# check if file or directory exists
|
364 |
+
if os.path.isfile(path) or os.path.islink(path):
|
365 |
+
# remove file
|
366 |
+
os.remove(path)
|
367 |
+
elif os.path.isdir(path):
|
368 |
+
# remove directory and all its content
|
369 |
+
shutil.rmtree(path)
|
370 |
+
else:
|
371 |
+
raise ValueError("Path {} is not a file or dir.".format(path))
|
372 |
+
|
373 |
+
|
374 |
+
def process_language_arg(language: str, model_name: str):
|
375 |
+
"""
|
376 |
+
Process the language argument to make sure it's valid and convert language names to language codes.
|
377 |
+
"""
|
378 |
+
if language is not None:
|
379 |
+
language = language.lower()
|
380 |
+
if language not in LANGUAGES:
|
381 |
+
if language in TO_LANGUAGE_CODE:
|
382 |
+
language = TO_LANGUAGE_CODE[language]
|
383 |
+
else:
|
384 |
+
raise ValueError(f"Unsupported language: {language}")
|
385 |
+
|
386 |
+
if model_name.endswith(".en") and language != "en":
|
387 |
+
if language is not None:
|
388 |
+
logging.warning(
|
389 |
+
f"{model_name} is an English-only model but received '{language}'; using English instead."
|
390 |
+
)
|
391 |
+
language = "en"
|
392 |
+
return language
|
mergefolders.py
ADDED
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
def list_directories(path):
|
4 |
+
"""List directories in the given path."""
|
5 |
+
return [d for d in os.listdir(path) if os.path.isdir(os.path.join(path, d))]
|
6 |
+
|
7 |
+
def select_directory(directories):
|
8 |
+
"""Let user select a directory from the list."""
|
9 |
+
for i, directory in enumerate(directories, start=1):
|
10 |
+
print(f"{i}. {directory}")
|
11 |
+
choice = int(input("Select a directory by number: ")) - 1
|
12 |
+
return directories[choice]
|
13 |
+
|
14 |
+
def merge_datasets(base_dir, dir1, dir2, output_dir):
|
15 |
+
"""Merge two LJ Speech datasets into one."""
|
16 |
+
ensure_dir(output_dir)
|
17 |
+
metadata_lines = []
|
18 |
+
|
19 |
+
for dir_name in [dir1, dir2]:
|
20 |
+
dir_path = os.path.join(base_dir, dir_name)
|
21 |
+
metadata_file = os.path.join(dir_path, "metadata.csv")
|
22 |
+
|
23 |
+
with open(metadata_file, "r", encoding="utf-8") as f:
|
24 |
+
lines = f.readlines()
|
25 |
+
for line in lines:
|
26 |
+
filename, transcription, normalized = line.strip().split("|")
|
27 |
+
# Copy audio file to the output directory
|
28 |
+
src_file_path = os.path.join(dir_path, filename + ".wav")
|
29 |
+
dst_file_path = os.path.join(output_dir, filename + ".wav")
|
30 |
+
os.system(f"cp '{src_file_path}' '{dst_file_path}'")
|
31 |
+
metadata_lines.append(line.strip())
|
32 |
+
|
33 |
+
# Save merged metadata
|
34 |
+
merged_metadata_file = os.path.join(output_dir, "metadata.csv")
|
35 |
+
with open(merged_metadata_file, "w", encoding="utf-8") as f:
|
36 |
+
f.write("\n".join(metadata_lines))
|
37 |
+
|
38 |
+
print(f"Merged dataset created in {output_dir}")
|
39 |
+
|
40 |
+
def ensure_dir(directory):
|
41 |
+
"""Ensure the directory exists."""
|
42 |
+
if not os.path.exists(directory):
|
43 |
+
os.makedirs(directory)
|
44 |
+
|
45 |
+
# Main process
|
46 |
+
base_dir = "LJ_Speech_dataset"
|
47 |
+
directories = list_directories(base_dir)
|
48 |
+
|
49 |
+
print("Select the first directory:")
|
50 |
+
first_dir = select_directory(directories)
|
51 |
+
|
52 |
+
print("Select the second directory:")
|
53 |
+
second_dir = select_directory(directories)
|
54 |
+
|
55 |
+
output_dir = os.path.join(base_dir, "Merged_Dataset")
|
56 |
+
merge_datasets(base_dir, first_dir, second_dir, output_dir)
|
nemo_msdd_configs/diar_infer_general.yaml
ADDED
@@ -0,0 +1,92 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# This YAML file is created for all types of offline speaker diarization inference tasks in `<NeMo git root>/example/speaker_tasks/diarization` folder.
|
2 |
+
# The inference parameters for VAD, speaker embedding extractor, clustering module, MSDD module, ASR decoder are all included in this YAML file.
|
3 |
+
# All the keys under `diarizer` key (`vad`, `speaker_embeddings`, `clustering`, `msdd_model`, `asr`) can be selectively used for its own purpose and also can be ignored if the module is not used.
|
4 |
+
# The configurations in this YAML file is optimized to show balanced performances on various types of domain. VAD is optimized on multilingual ASR datasets and diarizer is optimized on DIHARD3 development set.
|
5 |
+
# An example line in an input manifest file (`.json` format):
|
6 |
+
# {"audio_filepath": "/path/to/audio_file", "offset": 0, "duration": null, "label": "infer", "text": "-", "num_speakers": null, "rttm_filepath": "/path/to/rttm/file", "uem_filepath": "/path/to/uem/file"}
|
7 |
+
name: &name "ClusterDiarizer"
|
8 |
+
|
9 |
+
num_workers: 1
|
10 |
+
sample_rate: 16000
|
11 |
+
batch_size: 64
|
12 |
+
device: null # can specify a specific device, i.e: cuda:1 (default cuda if cuda available, else cpu)
|
13 |
+
verbose: True # enable additional logging
|
14 |
+
|
15 |
+
diarizer:
|
16 |
+
manifest_filepath: ???
|
17 |
+
out_dir: ???
|
18 |
+
oracle_vad: False # If True, uses RTTM files provided in the manifest file to get speech activity (VAD) timestamps
|
19 |
+
collar: 0.25 # Collar value for scoring
|
20 |
+
ignore_overlap: True # Consider or ignore overlap segments while scoring
|
21 |
+
|
22 |
+
vad:
|
23 |
+
model_path: vad_multilingual_marblenet # .nemo local model path or pretrained VAD model name
|
24 |
+
external_vad_manifest: null # This option is provided to use external vad and provide its speech activity labels for speaker embeddings extraction. Only one of model_path or external_vad_manifest should be set
|
25 |
+
|
26 |
+
parameters: # Tuned by detection error rate (false alarm + miss) on multilingual ASR evaluation datasets
|
27 |
+
window_length_in_sec: 0.63 # Window length in sec for VAD context input
|
28 |
+
shift_length_in_sec: 0.08 # Shift length in sec for generate frame level VAD prediction
|
29 |
+
smoothing: False # False or type of smoothing method (eg: median)
|
30 |
+
overlap: 0.5 # Overlap ratio for overlapped mean/median smoothing filter
|
31 |
+
onset: 0.5 # Onset threshold for detecting the beginning and end of a speech
|
32 |
+
offset: 0.3 # Offset threshold for detecting the end of a speech
|
33 |
+
pad_onset: 0.2 # Adding durations before each speech segment
|
34 |
+
pad_offset: 0.2 # Adding durations after each speech segment
|
35 |
+
min_duration_on: 0.5 # Threshold for small non_speech deletion
|
36 |
+
min_duration_off: 0.5 # Threshold for short speech segment deletion
|
37 |
+
filter_speech_first: True
|
38 |
+
|
39 |
+
speaker_embeddings:
|
40 |
+
model_path: titanet_large # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet)
|
41 |
+
parameters:
|
42 |
+
window_length_in_sec: [1.9,1.2,0.5] # Window length(s) in sec (floating-point number). either a number or a list. ex) 1.5 or [1.5,1.0,0.5]
|
43 |
+
shift_length_in_sec: [0.95,0.6,0.25] # Shift length(s) in sec (floating-point number). either a number or a list. ex) 0.75 or [0.75,0.5,0.25]
|
44 |
+
multiscale_weights: [1,1,1] # Weight for each scale. should be null (for single scale) or a list matched with window/shift scale count. ex) [0.33,0.33,0.33]
|
45 |
+
save_embeddings: True # If True, save speaker embeddings in pickle format. This should be True if clustering result is used for other models, such as `msdd_model`.
|
46 |
+
|
47 |
+
clustering:
|
48 |
+
parameters:
|
49 |
+
oracle_num_speakers: False # If True, use num of speakers value provided in manifest file.
|
50 |
+
max_num_speakers: 8 # Max number of speakers for each recording. If an oracle number of speakers is passed, this value is ignored.
|
51 |
+
enhanced_count_thres: 80 # If the number of segments is lower than this number, enhanced speaker counting is activated.
|
52 |
+
max_rp_threshold: 0.25 # Determines the range of p-value search: 0 < p <= max_rp_threshold.
|
53 |
+
sparse_search_volume: 10 # The higher the number, the more values will be examined with more time.
|
54 |
+
maj_vote_spk_count: False # If True, take a majority vote on multiple p-values to estimate the number of speakers.
|
55 |
+
|
56 |
+
msdd_model:
|
57 |
+
model_path: null # .nemo local model path or pretrained model name for multiscale diarization decoder (MSDD)
|
58 |
+
parameters:
|
59 |
+
use_speaker_model_from_ckpt: True # If True, use speaker embedding model in checkpoint. If False, the provided speaker embedding model in config will be used.
|
60 |
+
infer_batch_size: 25 # Batch size for MSDD inference.
|
61 |
+
sigmoid_threshold: [0.7] # Sigmoid threshold for generating binarized speaker labels. The smaller the more generous on detecting overlaps.
|
62 |
+
seq_eval_mode: False # If True, use oracle number of speaker and evaluate F1 score for the given speaker sequences. Default is False.
|
63 |
+
split_infer: True # If True, break the input audio clip to short sequences and calculate cluster average embeddings for inference.
|
64 |
+
diar_window_length: 50 # The length of split short sequence when split_infer is True.
|
65 |
+
overlap_infer_spk_limit: 5 # If the estimated number of speakers are larger than this number, overlap speech is not estimated.
|
66 |
+
|
67 |
+
asr:
|
68 |
+
model_path: null # Provide NGC cloud ASR model name. stt_en_conformer_ctc_* models are recommended for diarization purposes.
|
69 |
+
parameters:
|
70 |
+
asr_based_vad: False # if True, speech segmentation for diarization is based on word-timestamps from ASR inference.
|
71 |
+
asr_based_vad_threshold: 1.0 # Threshold (in sec) that caps the gap between two words when generating VAD timestamps using ASR based VAD.
|
72 |
+
asr_batch_size: null # Batch size can be dependent on each ASR model. Default batch sizes are applied if set to null.
|
73 |
+
decoder_delay_in_sec: null # Native decoder delay. null is recommended to use the default values for each ASR model.
|
74 |
+
word_ts_anchor_offset: null # Offset to set a reference point from the start of the word. Recommended range of values is [-0.05 0.2].
|
75 |
+
word_ts_anchor_pos: "start" # Select which part of the word timestamp we want to use. The options are: 'start', 'end', 'mid'.
|
76 |
+
fix_word_ts_with_VAD: False # Fix the word timestamp using VAD output. You must provide a VAD model to use this feature.
|
77 |
+
colored_text: False # If True, use colored text to distinguish speakers in the output transcript.
|
78 |
+
print_time: True # If True, the start and end time of each speaker turn is printed in the output transcript.
|
79 |
+
break_lines: False # If True, the output transcript breaks the line to fix the line width (default is 90 chars)
|
80 |
+
|
81 |
+
ctc_decoder_parameters: # Optional beam search decoder (pyctcdecode)
|
82 |
+
pretrained_language_model: null # KenLM model file: .arpa model file or .bin binary file.
|
83 |
+
beam_width: 32
|
84 |
+
alpha: 0.5
|
85 |
+
beta: 2.5
|
86 |
+
|
87 |
+
realigning_lm_parameters: # Experimental feature
|
88 |
+
arpa_language_model: null # Provide a KenLM language model in .arpa format.
|
89 |
+
min_number_of_words: 3 # Min number of words for the left context.
|
90 |
+
max_number_of_words: 10 # Max number of words for the right context.
|
91 |
+
logprob_diff_threshold: 1.2 # The threshold for the difference between two log probability values from two hypotheses.
|
92 |
+
|
nemo_msdd_configs/diar_infer_meeting.yaml
ADDED
@@ -0,0 +1,92 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# This YAML file is created for all types of offline speaker diarization inference tasks in `<NeMo git root>/example/speaker_tasks/diarization` folder.
|
2 |
+
# The inference parameters for VAD, speaker embedding extractor, clustering module, MSDD module, ASR decoder are all included in this YAML file.
|
3 |
+
# All the keys under `diarizer` key (`vad`, `speaker_embeddings`, `clustering`, `msdd_model`, `asr`) can be selectively used for its own purpose and also can be ignored if the module is not used.
|
4 |
+
# The configurations in this YAML file is suitable for 3~5 speakers participating in a meeting and may not show the best performance on other types of dialogues.
|
5 |
+
# An example line in an input manifest file (`.json` format):
|
6 |
+
# {"audio_filepath": "/path/to/audio_file", "offset": 0, "duration": null, "label": "infer", "text": "-", "num_speakers": null, "rttm_filepath": "/path/to/rttm/file", "uem_filepath": "/path/to/uem/file"}
|
7 |
+
name: &name "ClusterDiarizer"
|
8 |
+
|
9 |
+
num_workers: 1
|
10 |
+
sample_rate: 16000
|
11 |
+
batch_size: 64
|
12 |
+
device: null # can specify a specific device, i.e: cuda:1 (default cuda if cuda available, else cpu)
|
13 |
+
verbose: True # enable additional logging
|
14 |
+
|
15 |
+
diarizer:
|
16 |
+
manifest_filepath: ???
|
17 |
+
out_dir: ???
|
18 |
+
oracle_vad: False # If True, uses RTTM files provided in the manifest file to get speech activity (VAD) timestamps
|
19 |
+
collar: 0.25 # Collar value for scoring
|
20 |
+
ignore_overlap: True # Consider or ignore overlap segments while scoring
|
21 |
+
|
22 |
+
vad:
|
23 |
+
model_path: vad_multilingual_marblenet # .nemo local model path or pretrained VAD model name
|
24 |
+
external_vad_manifest: null # This option is provided to use external vad and provide its speech activity labels for speaker embeddings extraction. Only one of model_path or external_vad_manifest should be set
|
25 |
+
|
26 |
+
parameters: # Tuned parameters for CH109 (using the 11 multi-speaker sessions as dev set)
|
27 |
+
window_length_in_sec: 0.63 # Window length in sec for VAD context input
|
28 |
+
shift_length_in_sec: 0.01 # Shift length in sec for generate frame level VAD prediction
|
29 |
+
smoothing: False # False or type of smoothing method (eg: median)
|
30 |
+
overlap: 0.5 # Overlap ratio for overlapped mean/median smoothing filter
|
31 |
+
onset: 0.9 # Onset threshold for detecting the beginning and end of a speech
|
32 |
+
offset: 0.5 # Offset threshold for detecting the end of a speech
|
33 |
+
pad_onset: 0 # Adding durations before each speech segment
|
34 |
+
pad_offset: 0 # Adding durations after each speech segment
|
35 |
+
min_duration_on: 0 # Threshold for small non_speech deletion
|
36 |
+
min_duration_off: 0.6 # Threshold for short speech segment deletion
|
37 |
+
filter_speech_first: True
|
38 |
+
|
39 |
+
speaker_embeddings:
|
40 |
+
model_path: titanet_large # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet)
|
41 |
+
parameters:
|
42 |
+
window_length_in_sec: [3.0,2.5,2.0,1.5,1.0,0.5] # Window length(s) in sec (floating-point number). either a number or a list. ex) 1.5 or [1.5,1.0,0.5]
|
43 |
+
shift_length_in_sec: [1.5,1.25,1.0,0.75,0.5,0.25] # Shift length(s) in sec (floating-point number). either a number or a list. ex) 0.75 or [0.75,0.5,0.25]
|
44 |
+
multiscale_weights: [1,1,1,1,1,1] # Weight for each scale. should be null (for single scale) or a list matched with window/shift scale count. ex) [0.33,0.33,0.33]
|
45 |
+
save_embeddings: True # If True, save speaker embeddings in pickle format. This should be True if clustering result is used for other models, such as `msdd_model`.
|
46 |
+
|
47 |
+
clustering:
|
48 |
+
parameters:
|
49 |
+
oracle_num_speakers: False # If True, use num of speakers value provided in manifest file.
|
50 |
+
max_num_speakers: 8 # Max number of speakers for each recording. If an oracle number of speakers is passed, this value is ignored.
|
51 |
+
enhanced_count_thres: 80 # If the number of segments is lower than this number, enhanced speaker counting is activated.
|
52 |
+
max_rp_threshold: 0.25 # Determines the range of p-value search: 0 < p <= max_rp_threshold.
|
53 |
+
sparse_search_volume: 30 # The higher the number, the more values will be examined with more time.
|
54 |
+
maj_vote_spk_count: False # If True, take a majority vote on multiple p-values to estimate the number of speakers.
|
55 |
+
|
56 |
+
msdd_model:
|
57 |
+
model_path: null # .nemo local model path or pretrained model name for multiscale diarization decoder (MSDD)
|
58 |
+
parameters:
|
59 |
+
use_speaker_model_from_ckpt: True # If True, use speaker embedding model in checkpoint. If False, the provided speaker embedding model in config will be used.
|
60 |
+
infer_batch_size: 25 # Batch size for MSDD inference.
|
61 |
+
sigmoid_threshold: [0.7] # Sigmoid threshold for generating binarized speaker labels. The smaller the more generous on detecting overlaps.
|
62 |
+
seq_eval_mode: False # If True, use oracle number of speaker and evaluate F1 score for the given speaker sequences. Default is False.
|
63 |
+
split_infer: True # If True, break the input audio clip to short sequences and calculate cluster average embeddings for inference.
|
64 |
+
diar_window_length: 50 # The length of split short sequence when split_infer is True.
|
65 |
+
overlap_infer_spk_limit: 5 # If the estimated number of speakers are larger than this number, overlap speech is not estimated.
|
66 |
+
|
67 |
+
asr:
|
68 |
+
model_path: stt_en_conformer_ctc_large # Provide NGC cloud ASR model name. stt_en_conformer_ctc_* models are recommended for diarization purposes.
|
69 |
+
parameters:
|
70 |
+
asr_based_vad: False # if True, speech segmentation for diarization is based on word-timestamps from ASR inference.
|
71 |
+
asr_based_vad_threshold: 1.0 # Threshold (in sec) that caps the gap between two words when generating VAD timestamps using ASR based VAD.
|
72 |
+
asr_batch_size: null # Batch size can be dependent on each ASR model. Default batch sizes are applied if set to null.
|
73 |
+
decoder_delay_in_sec: null # Native decoder delay. null is recommended to use the default values for each ASR model.
|
74 |
+
word_ts_anchor_offset: null # Offset to set a reference point from the start of the word. Recommended range of values is [-0.05 0.2].
|
75 |
+
word_ts_anchor_pos: "start" # Select which part of the word timestamp we want to use. The options are: 'start', 'end', 'mid'.
|
76 |
+
fix_word_ts_with_VAD: False # Fix the word timestamp using VAD output. You must provide a VAD model to use this feature.
|
77 |
+
colored_text: False # If True, use colored text to distinguish speakers in the output transcript.
|
78 |
+
print_time: True # If True, the start and end time of each speaker turn is printed in the output transcript.
|
79 |
+
break_lines: False # If True, the output transcript breaks the line to fix the line width (default is 90 chars)
|
80 |
+
|
81 |
+
ctc_decoder_parameters: # Optional beam search decoder (pyctcdecode)
|
82 |
+
pretrained_language_model: null # KenLM model file: .arpa model file or .bin binary file.
|
83 |
+
beam_width: 32
|
84 |
+
alpha: 0.5
|
85 |
+
beta: 2.5
|
86 |
+
|
87 |
+
realigning_lm_parameters: # Experimental feature
|
88 |
+
arpa_language_model: null # Provide a KenLM language model in .arpa format.
|
89 |
+
min_number_of_words: 3 # Min number of words for the left context.
|
90 |
+
max_number_of_words: 10 # Max number of words for the right context.
|
91 |
+
logprob_diff_threshold: 1.2 # The threshold for the difference between two log probability values from two hypotheses.
|
92 |
+
|
nemo_msdd_configs/diar_infer_telephonic.yaml
ADDED
@@ -0,0 +1,92 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# This YAML file is created for all types of offline speaker diarization inference tasks in `<NeMo git root>/example/speaker_tasks/diarization` folder.
|
2 |
+
# The inference parameters for VAD, speaker embedding extractor, clustering module, MSDD module, ASR decoder are all included in this YAML file.
|
3 |
+
# All the keys under `diarizer` key (`vad`, `speaker_embeddings`, `clustering`, `msdd_model`, `asr`) can be selectively used for its own purpose and also can be ignored if the module is not used.
|
4 |
+
# The configurations in this YAML file is suitable for telephone recordings involving 2~8 speakers in a session and may not show the best performance on the other types of acoustic conditions or dialogues.
|
5 |
+
# An example line in an input manifest file (`.json` format):
|
6 |
+
# {"audio_filepath": "/path/to/audio_file", "offset": 0, "duration": null, "label": "infer", "text": "-", "num_speakers": null, "rttm_filepath": "/path/to/rttm/file", "uem_filepath": "/path/to/uem/file"}
|
7 |
+
name: &name "ClusterDiarizer"
|
8 |
+
|
9 |
+
num_workers: 1
|
10 |
+
sample_rate: 16000
|
11 |
+
batch_size: 64
|
12 |
+
device: null # can specify a specific device, i.e: cuda:1 (default cuda if cuda available, else cpu)
|
13 |
+
verbose: True # enable additional logging
|
14 |
+
|
15 |
+
diarizer:
|
16 |
+
manifest_filepath: ???
|
17 |
+
out_dir: ???
|
18 |
+
oracle_vad: False # If True, uses RTTM files provided in the manifest file to get speech activity (VAD) timestamps
|
19 |
+
collar: 0.25 # Collar value for scoring
|
20 |
+
ignore_overlap: True # Consider or ignore overlap segments while scoring
|
21 |
+
|
22 |
+
vad:
|
23 |
+
model_path: vad_multilingual_marblenet # .nemo local model path or pretrained VAD model name
|
24 |
+
external_vad_manifest: null # This option is provided to use external vad and provide its speech activity labels for speaker embeddings extraction. Only one of model_path or external_vad_manifest should be set
|
25 |
+
|
26 |
+
parameters: # Tuned parameters for CH109 (using the 11 multi-speaker sessions as dev set)
|
27 |
+
window_length_in_sec: 0.15 # Window length in sec for VAD context input
|
28 |
+
shift_length_in_sec: 0.01 # Shift length in sec for generate frame level VAD prediction
|
29 |
+
smoothing: "median" # False or type of smoothing method (eg: median)
|
30 |
+
overlap: 0.5 # Overlap ratio for overlapped mean/median smoothing filter
|
31 |
+
onset: 0.1 # Onset threshold for detecting the beginning and end of a speech
|
32 |
+
offset: 0.1 # Offset threshold for detecting the end of a speech
|
33 |
+
pad_onset: 0.1 # Adding durations before each speech segment
|
34 |
+
pad_offset: 0 # Adding durations after each speech segment
|
35 |
+
min_duration_on: 0 # Threshold for small non_speech deletion
|
36 |
+
min_duration_off: 0.2 # Threshold for short speech segment deletion
|
37 |
+
filter_speech_first: True
|
38 |
+
|
39 |
+
speaker_embeddings:
|
40 |
+
model_path: titanet_large # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet)
|
41 |
+
parameters:
|
42 |
+
window_length_in_sec: [1.5,1.25,1.0,0.75,0.5] # Window length(s) in sec (floating-point number). either a number or a list. ex) 1.5 or [1.5,1.0,0.5]
|
43 |
+
shift_length_in_sec: [0.75,0.625,0.5,0.375,0.25] # Shift length(s) in sec (floating-point number). either a number or a list. ex) 0.75 or [0.75,0.5,0.25]
|
44 |
+
multiscale_weights: [1,1,1,1,1] # Weight for each scale. should be null (for single scale) or a list matched with window/shift scale count. ex) [0.33,0.33,0.33]
|
45 |
+
save_embeddings: True # If True, save speaker embeddings in pickle format. This should be True if clustering result is used for other models, such as `msdd_model`.
|
46 |
+
|
47 |
+
clustering:
|
48 |
+
parameters:
|
49 |
+
oracle_num_speakers: False # If True, use num of speakers value provided in manifest file.
|
50 |
+
max_num_speakers: 8 # Max number of speakers for each recording. If an oracle number of speakers is passed, this value is ignored.
|
51 |
+
enhanced_count_thres: 80 # If the number of segments is lower than this number, enhanced speaker counting is activated.
|
52 |
+
max_rp_threshold: 0.25 # Determines the range of p-value search: 0 < p <= max_rp_threshold.
|
53 |
+
sparse_search_volume: 30 # The higher the number, the more values will be examined with more time.
|
54 |
+
maj_vote_spk_count: False # If True, take a majority vote on multiple p-values to estimate the number of speakers.
|
55 |
+
|
56 |
+
msdd_model:
|
57 |
+
model_path: diar_msdd_telephonic # .nemo local model path or pretrained model name for multiscale diarization decoder (MSDD)
|
58 |
+
parameters:
|
59 |
+
use_speaker_model_from_ckpt: True # If True, use speaker embedding model in checkpoint. If False, the provided speaker embedding model in config will be used.
|
60 |
+
infer_batch_size: 25 # Batch size for MSDD inference.
|
61 |
+
sigmoid_threshold: [0.7] # Sigmoid threshold for generating binarized speaker labels. The smaller the more generous on detecting overlaps.
|
62 |
+
seq_eval_mode: False # If True, use oracle number of speaker and evaluate F1 score for the given speaker sequences. Default is False.
|
63 |
+
split_infer: True # If True, break the input audio clip to short sequences and calculate cluster average embeddings for inference.
|
64 |
+
diar_window_length: 50 # The length of split short sequence when split_infer is True.
|
65 |
+
overlap_infer_spk_limit: 5 # If the estimated number of speakers are larger than this number, overlap speech is not estimated.
|
66 |
+
|
67 |
+
asr:
|
68 |
+
model_path: stt_en_conformer_ctc_large # Provide NGC cloud ASR model name. stt_en_conformer_ctc_* models are recommended for diarization purposes.
|
69 |
+
parameters:
|
70 |
+
asr_based_vad: False # if True, speech segmentation for diarization is based on word-timestamps from ASR inference.
|
71 |
+
asr_based_vad_threshold: 1.0 # Threshold (in sec) that caps the gap between two words when generating VAD timestamps using ASR based VAD.
|
72 |
+
asr_batch_size: null # Batch size can be dependent on each ASR model. Default batch sizes are applied if set to null.
|
73 |
+
decoder_delay_in_sec: null # Native decoder delay. null is recommended to use the default values for each ASR model.
|
74 |
+
word_ts_anchor_offset: null # Offset to set a reference point from the start of the word. Recommended range of values is [-0.05 0.2].
|
75 |
+
word_ts_anchor_pos: "start" # Select which part of the word timestamp we want to use. The options are: 'start', 'end', 'mid'.
|
76 |
+
fix_word_ts_with_VAD: False # Fix the word timestamp using VAD output. You must provide a VAD model to use this feature.
|
77 |
+
colored_text: False # If True, use colored text to distinguish speakers in the output transcript.
|
78 |
+
print_time: True # If True, the start and end time of each speaker turn is printed in the output transcript.
|
79 |
+
break_lines: False # If True, the output transcript breaks the line to fix the line width (default is 90 chars)
|
80 |
+
|
81 |
+
ctc_decoder_parameters: # Optional beam search decoder (pyctcdecode)
|
82 |
+
pretrained_language_model: null # KenLM model file: .arpa model file or .bin binary file.
|
83 |
+
beam_width: 32
|
84 |
+
alpha: 0.5
|
85 |
+
beta: 2.5
|
86 |
+
|
87 |
+
realigning_lm_parameters: # Experimental feature
|
88 |
+
arpa_language_model: null # Provide a KenLM language model in .arpa format.
|
89 |
+
min_number_of_words: 3 # Min number of words for the left context.
|
90 |
+
max_number_of_words: 10 # Max number of words for the right context.
|
91 |
+
logprob_diff_threshold: 1.2 # The threshold for the difference between two log probability values from two hypotheses.
|
92 |
+
|
nemo_process.py
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
from helpers import *
|
4 |
+
import torch
|
5 |
+
from pydub import AudioSegment
|
6 |
+
from nemo.collections.asr.models.msdd_models import NeuralDiarizer
|
7 |
+
|
8 |
+
parser = argparse.ArgumentParser()
|
9 |
+
parser.add_argument(
|
10 |
+
"-a", "--audio", help="name of the target audio file", required=True
|
11 |
+
)
|
12 |
+
parser.add_argument(
|
13 |
+
"--device",
|
14 |
+
dest="device",
|
15 |
+
default="cuda" if torch.cuda.is_available() else "cpu",
|
16 |
+
help="if you have a GPU use 'cuda', otherwise 'cpu'",
|
17 |
+
)
|
18 |
+
args = parser.parse_args()
|
19 |
+
|
20 |
+
# convert audio to mono for NeMo combatibility
|
21 |
+
sound = AudioSegment.from_file(args.audio).set_channels(1)
|
22 |
+
ROOT = os.getcwd()
|
23 |
+
temp_path = os.path.join(ROOT, "temp_outputs")
|
24 |
+
os.makedirs(temp_path, exist_ok=True)
|
25 |
+
sound.export(os.path.join(temp_path, "mono_file.wav"), format="wav")
|
26 |
+
|
27 |
+
# Initialize NeMo MSDD diarization model
|
28 |
+
msdd_model = NeuralDiarizer(cfg=create_config(temp_path)).to(args.device)
|
29 |
+
msdd_model.diarize()
|
process_srt_wav.py
ADDED
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pysrt
|
2 |
+
import os
|
3 |
+
from pydub import AudioSegment
|
4 |
+
|
5 |
+
# Function to ensure directory exists
|
6 |
+
def ensure_dir(directory):
|
7 |
+
if not os.path.exists(directory):
|
8 |
+
os.makedirs(directory)
|
9 |
+
|
10 |
+
# Function to find the first unique SRT and WAV combo
|
11 |
+
def find_unique_combo():
|
12 |
+
for file in os.listdir():
|
13 |
+
if file.endswith(".srt"):
|
14 |
+
srt_file = file
|
15 |
+
wav_file = file[:-4] + ".wav"
|
16 |
+
if os.path.exists(wav_file):
|
17 |
+
return srt_file, wav_file
|
18 |
+
return None, None
|
19 |
+
|
20 |
+
# Find the first unique SRT and WAV combo
|
21 |
+
srt_file, wav_file = find_unique_combo()
|
22 |
+
|
23 |
+
if srt_file and wav_file:
|
24 |
+
# Load the SRT file
|
25 |
+
subs = pysrt.open(srt_file)
|
26 |
+
# Load the WAV file
|
27 |
+
audio = AudioSegment.from_wav(wav_file)
|
28 |
+
|
29 |
+
# Base directory for the LJ Speech-like structure
|
30 |
+
base_dir = "LJ_Speech_dataset"
|
31 |
+
# Dictionary to hold audio segments and texts for each speaker
|
32 |
+
speaker_audios_texts = {}
|
33 |
+
|
34 |
+
# Process each subtitle
|
35 |
+
for sub in subs:
|
36 |
+
start_time = (sub.start.hours * 3600 + sub.start.minutes * 60 + sub.start.seconds) * 1000 + sub.start.milliseconds
|
37 |
+
end_time = (sub.end.hours * 3600 + sub.end.minutes * 60 + sub.end.seconds) * 1000 + sub.end.milliseconds
|
38 |
+
|
39 |
+
# Extract speaker and text from the subtitle
|
40 |
+
speaker_text = sub.text.split(':')
|
41 |
+
if len(speaker_text) > 1:
|
42 |
+
speaker = speaker_text[0].strip()
|
43 |
+
text = ':'.join(speaker_text[1:]).strip()
|
44 |
+
segment = audio[start_time:end_time]
|
45 |
+
|
46 |
+
# Append or create the audio segment and text for the speaker
|
47 |
+
if speaker not in speaker_audios_texts:
|
48 |
+
speaker_audios_texts[speaker] = []
|
49 |
+
speaker_audios_texts[speaker].append((segment, text))
|
50 |
+
|
51 |
+
# Save each speaker's audio to a separate file and generate metadata
|
52 |
+
for speaker, segments_texts in speaker_audios_texts.items():
|
53 |
+
speaker_dir = os.path.join(base_dir, speaker.replace(' ', '_'))
|
54 |
+
ensure_dir(speaker_dir)
|
55 |
+
|
56 |
+
metadata_lines = []
|
57 |
+
for i, (segment, text) in enumerate(segments_texts, start=1):
|
58 |
+
filename = f"{speaker.replace(' ', '_')}_{i:03}.wav"
|
59 |
+
filepath = os.path.join(speaker_dir, filename)
|
60 |
+
segment.export(filepath, format="wav")
|
61 |
+
|
62 |
+
# Prepare metadata line (filename without extension, speaker, text)
|
63 |
+
metadata_lines.append(f"{filename[:-4]}|{speaker}|{text}")
|
64 |
+
|
65 |
+
# Save metadata to a file
|
66 |
+
metadata_file = os.path.join(speaker_dir, "metadata.csv")
|
67 |
+
with open(metadata_file, "w", encoding="utf-8") as f:
|
68 |
+
f.write("\n".join(metadata_lines))
|
69 |
+
|
70 |
+
print(f"Exported files and metadata for {speaker}")
|
71 |
+
|
72 |
+
# Move the original WAV and SRT files to the "handled" subfolder
|
73 |
+
handled_dir = "handled"
|
74 |
+
ensure_dir(handled_dir)
|
75 |
+
os.rename(srt_file, os.path.join(handled_dir, srt_file))
|
76 |
+
os.rename(wav_file, os.path.join(handled_dir, wav_file))
|
77 |
+
|
78 |
+
print(f"Moved {srt_file} and {wav_file} to the 'handled' subfolder.")
|
79 |
+
else:
|
80 |
+
print("No unique SRT and WAV combo")
|
test.py
ADDED
@@ -0,0 +1,269 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import time
|
2 |
+
from datasets import Dataset
|
3 |
+
import warnings
|
4 |
+
import argparse
|
5 |
+
import os
|
6 |
+
from helpers import *
|
7 |
+
from faster_whisper import WhisperModel
|
8 |
+
import whisperx
|
9 |
+
import torch
|
10 |
+
from pydub import AudioSegment
|
11 |
+
from nemo.collections.asr.models.msdd_models import NeuralDiarizer
|
12 |
+
from deepmultilingualpunctuation import PunctuationModel
|
13 |
+
import re
|
14 |
+
import logging
|
15 |
+
import csv
|
16 |
+
import shutil
|
17 |
+
|
18 |
+
mtypes = {"cpu": "int8", "cuda": "float16"}
|
19 |
+
|
20 |
+
# Initialize parser
|
21 |
+
parser = argparse.ArgumentParser()
|
22 |
+
parser.add_argument(
|
23 |
+
"-a", "--audio", help="name of the target audio file", required=True
|
24 |
+
)
|
25 |
+
parser.add_argument(
|
26 |
+
"--no-stem",
|
27 |
+
action="store_false",
|
28 |
+
dest="stemming",
|
29 |
+
default=True,
|
30 |
+
help="Disables source separation."
|
31 |
+
"This helps with long files that don't contain a lot of music.",
|
32 |
+
)
|
33 |
+
parser.add_argument(
|
34 |
+
"--suppress_numerals",
|
35 |
+
action="store_true",
|
36 |
+
dest="suppress_numerals",
|
37 |
+
default=False,
|
38 |
+
help="Suppresses Numerical Digits."
|
39 |
+
"This helps the diarization accuracy but converts all digits into written text.",
|
40 |
+
)
|
41 |
+
parser.add_argument(
|
42 |
+
"--whisper-model",
|
43 |
+
dest="model_name",
|
44 |
+
default="medium.en",
|
45 |
+
help="name of the Whisper model to use",
|
46 |
+
)
|
47 |
+
parser.add_argument(
|
48 |
+
"--batch-size",
|
49 |
+
type=int,
|
50 |
+
dest="batch_size",
|
51 |
+
default=8,
|
52 |
+
help="Batch size for batched inference, reduce if you run out of memory, set to 0 for non-batched inference",
|
53 |
+
)
|
54 |
+
parser.add_argument(
|
55 |
+
"--language",
|
56 |
+
type=str,
|
57 |
+
default=None,
|
58 |
+
choices=whisper_langs,
|
59 |
+
help="Language spoken in the audio, specify None to perform language detection",
|
60 |
+
)
|
61 |
+
parser.add_argument(
|
62 |
+
"--device",
|
63 |
+
dest="device",
|
64 |
+
default="cuda" if torch.cuda.is_available() else "cpu",
|
65 |
+
help="if you have a GPU use 'cuda', otherwise 'cpu'",
|
66 |
+
)
|
67 |
+
args = parser.parse_args()
|
68 |
+
|
69 |
+
if args.stemming:
|
70 |
+
# Isolate vocals from the rest of the audio
|
71 |
+
return_code = os.system(
|
72 |
+
f'python3 -m demucs.separate -n htdemucs --two-stems=vocals "{args.audio}" -o "temp_outputs"'
|
73 |
+
)
|
74 |
+
if return_code != 0:
|
75 |
+
logging.warning(
|
76 |
+
"Source splitting failed, using original audio file. Use --no-stem argument to disable it."
|
77 |
+
)
|
78 |
+
vocal_target = args.audio
|
79 |
+
else:
|
80 |
+
vocal_target = os.path.join(
|
81 |
+
"temp_outputs",
|
82 |
+
"htdemucs",
|
83 |
+
os.path.splitext(os.path.basename(args.audio))[0],
|
84 |
+
"vocals.wav",
|
85 |
+
)
|
86 |
+
else:
|
87 |
+
vocal_target = args.audio
|
88 |
+
|
89 |
+
# Transcribe the audio file
|
90 |
+
if args.batch_size != 0:
|
91 |
+
from transcription_helpers import transcribe_batched
|
92 |
+
whisper_results, language = transcribe_batched(
|
93 |
+
vocal_target,
|
94 |
+
args.language,
|
95 |
+
args.batch_size,
|
96 |
+
args.model_name,
|
97 |
+
mtypes[args.device],
|
98 |
+
args.suppress_numerals,
|
99 |
+
args.device,
|
100 |
+
)
|
101 |
+
else:
|
102 |
+
from transcription_helpers import transcribe
|
103 |
+
whisper_results, language = transcribe(
|
104 |
+
vocal_target,
|
105 |
+
args.language,
|
106 |
+
args.model_name,
|
107 |
+
mtypes[args.device],
|
108 |
+
args.suppress_numerals,
|
109 |
+
args.device,
|
110 |
+
)
|
111 |
+
|
112 |
+
if language in wav2vec2_langs:
|
113 |
+
alignment_model, metadata = whisperx.load_align_model(
|
114 |
+
language_code=language, device=args.device
|
115 |
+
)
|
116 |
+
result_aligned = whisperx.align(
|
117 |
+
whisper_results, alignment_model, metadata, vocal_target, args.device
|
118 |
+
)
|
119 |
+
word_timestamps = filter_missing_timestamps(
|
120 |
+
result_aligned["word_segments"],
|
121 |
+
initial_timestamp=whisper_results[0].get("start"),
|
122 |
+
final_timestamp=whisper_results[-1].get("end"),
|
123 |
+
)
|
124 |
+
# clear gpu vram
|
125 |
+
del alignment_model
|
126 |
+
torch.cuda.empty_cache()
|
127 |
+
else:
|
128 |
+
assert (
|
129 |
+
args.batch_size == 0 # TODO: add a better check for word timestamps existence
|
130 |
+
), (
|
131 |
+
f"Unsupported language: {language}, use --batch_size to 0"
|
132 |
+
" to generate word timestamps using whisper directly and fix this error."
|
133 |
+
)
|
134 |
+
word_timestamps = []
|
135 |
+
for segment in whisper_results:
|
136 |
+
for word in segment["words"]:
|
137 |
+
word_timestamps.append({"word": word[2], "start": word[0], "end": word[1]})
|
138 |
+
|
139 |
+
# convert audio to mono for NeMo combatibility
|
140 |
+
sound = AudioSegment.from_file(vocal_target).set_channels(1)
|
141 |
+
ROOT = os.getcwd()
|
142 |
+
temp_path = os.path.join(ROOT, "temp_outputs")
|
143 |
+
os.makedirs(temp_path, exist_ok=True)
|
144 |
+
sound.export(os.path.join(temp_path, "mono_file.wav"), format="wav")
|
145 |
+
|
146 |
+
# Initialize NeMo MSDD diarization model
|
147 |
+
msdd_model = NeuralDiarizer(cfg=create_config(temp_path)).to(args.device)
|
148 |
+
msdd_model.diarize()
|
149 |
+
del msdd_model
|
150 |
+
torch.cuda.empty_cache()
|
151 |
+
|
152 |
+
# Reading timestamps <> Speaker Labels mapping
|
153 |
+
speaker_ts = []
|
154 |
+
with open(os.path.join(temp_path, "pred_rttms", "mono_file.rttm"), "r") as f:
|
155 |
+
lines = f.readlines()
|
156 |
+
for line in lines:
|
157 |
+
line_list = line.split(" ")
|
158 |
+
s = int(float(line_list[5]) * 1000)
|
159 |
+
e = s + int(float(line_list[8]) * 1000)
|
160 |
+
speaker_ts.append([s, e, int(line_list[11].split("_")[-1])])
|
161 |
+
|
162 |
+
wsm = get_words_speaker_mapping(word_timestamps, speaker_ts, "start")
|
163 |
+
|
164 |
+
if language in punct_model_langs:
|
165 |
+
# restoring punctuation in the transcript to help realign the sentences
|
166 |
+
punct_model = PunctuationModel(model="kredor/punctuate-all")
|
167 |
+
words_list = list(map(lambda x: x["word"], wsm))
|
168 |
+
|
169 |
+
# Use the pipe method directly on the words_list
|
170 |
+
while True:
|
171 |
+
try:
|
172 |
+
labled_words = punct_model.pipe(words_list)
|
173 |
+
break
|
174 |
+
except ValueError as e:
|
175 |
+
if str(e) == "Queue is full! Please try again.":
|
176 |
+
print("Queue is full. Retrying in 1 second...")
|
177 |
+
time.sleep(1)
|
178 |
+
else:
|
179 |
+
raise e
|
180 |
+
|
181 |
+
ending_puncts = ".?!"
|
182 |
+
model_puncts = ".,;:!?"
|
183 |
+
# We don't want to punctuate U.S.A. with a period. Right?
|
184 |
+
is_acronym = lambda x: re.fullmatch(r"\b(?:[a-zA-Z]\.){2,}", x)
|
185 |
+
for i, labeled_tuple in enumerate(labled_words):
|
186 |
+
word = wsm[i]["word"]
|
187 |
+
if (
|
188 |
+
word
|
189 |
+
and labeled_tuple
|
190 |
+
and "entity" in labeled_tuple[0]
|
191 |
+
and labeled_tuple[0]["entity"] in ending_puncts
|
192 |
+
and (word[-1] not in model_puncts or is_acronym(word))
|
193 |
+
):
|
194 |
+
word += labeled_tuple[0]["entity"]
|
195 |
+
if word.endswith(".."):
|
196 |
+
word = word.rstrip(".")
|
197 |
+
wsm[i]["word"] = word
|
198 |
+
else:
|
199 |
+
logging.warning(
|
200 |
+
f"Punctuation restoration is not available for {language} language. Using the original punctuation."
|
201 |
+
)
|
202 |
+
|
203 |
+
wsm = get_realigned_ws_mapping_with_punctuation(wsm)
|
204 |
+
ssm = get_sentences_speaker_mapping(wsm, speaker_ts)
|
205 |
+
|
206 |
+
|
207 |
+
with open(f"{os.path.splitext(args.audio)[0]}.txt", "w", encoding="utf-8-sig") as f:
|
208 |
+
get_speaker_aware_transcript(ssm, f)
|
209 |
+
with open(f"{os.path.splitext(args.audio)[0]}.srt", "w", encoding="utf-8-sig") as srt_file:
|
210 |
+
write_srt(ssm, srt_file)
|
211 |
+
|
212 |
+
# Create the autodiarization directory structure
|
213 |
+
autodiarization_dir = "autodiarization"
|
214 |
+
os.makedirs(autodiarization_dir, exist_ok=True)
|
215 |
+
|
216 |
+
# Get the base name of the audio file
|
217 |
+
audio_base_name = os.path.splitext(os.path.basename(args.audio))[0]
|
218 |
+
|
219 |
+
# Determine the next available subdirectory number
|
220 |
+
subdirs = [int(d) for d in os.listdir(autodiarization_dir) if os.path.isdir(os.path.join(autodiarization_dir, d))]
|
221 |
+
next_subdir = str(max(subdirs) + 1) if subdirs else "0"
|
222 |
+
|
223 |
+
# Create the subdirectory for the current audio file
|
224 |
+
audio_subdir = os.path.join(autodiarization_dir, next_subdir)
|
225 |
+
os.makedirs(audio_subdir, exist_ok=True)
|
226 |
+
|
227 |
+
# Read the SRT file
|
228 |
+
with open(f"{os.path.splitext(args.audio)[0]}.srt", "r", encoding="utf-8-sig") as srt_file:
|
229 |
+
srt_data = srt_file.read()
|
230 |
+
|
231 |
+
# Parse the SRT data
|
232 |
+
srt_parser = srt.parse(srt_data)
|
233 |
+
|
234 |
+
# Split the audio file based on the SRT timestamps and create the LJSpeech dataset
|
235 |
+
speaker_dirs = {}
|
236 |
+
for index, subtitle in enumerate(srt_parser):
|
237 |
+
start_time = subtitle.start.total_seconds()
|
238 |
+
end_time = subtitle.end.total_seconds()
|
239 |
+
|
240 |
+
# Extract the speaker information from the TXT file
|
241 |
+
with open(f"{os.path.splitext(args.audio)[0]}.txt", "r", encoding="utf-8-sig") as txt_file:
|
242 |
+
for line in txt_file:
|
243 |
+
if f"{index+1}" in line:
|
244 |
+
speaker = line.split(":")[0].strip()
|
245 |
+
break
|
246 |
+
|
247 |
+
if speaker not in speaker_dirs:
|
248 |
+
speaker_dir = os.path.join(audio_subdir, speaker)
|
249 |
+
os.makedirs(speaker_dir, exist_ok=True)
|
250 |
+
speaker_dirs[speaker] = speaker_dir
|
251 |
+
|
252 |
+
# Extract the audio segment for the current subtitle
|
253 |
+
audio_segment = sound[start_time * 1000:end_time * 1000]
|
254 |
+
|
255 |
+
# Generate a unique filename for the audio segment
|
256 |
+
segment_filename = f"{speaker}_{len(os.listdir(speaker_dirs[speaker])) + 1:03d}.wav"
|
257 |
+
segment_path = os.path.join(speaker_dirs[speaker], segment_filename)
|
258 |
+
|
259 |
+
# Export the audio segment as a WAV file
|
260 |
+
audio_segment.export(segment_path, format="wav")
|
261 |
+
|
262 |
+
# Append the metadata to the CSV file
|
263 |
+
metadata_path = os.path.join(speaker_dirs[speaker], "metadata.csv")
|
264 |
+
with open(metadata_path, "a", newline="", encoding="utf-8-sig") as csvfile:
|
265 |
+
writer = csv.writer(csvfile, delimiter="|")
|
266 |
+
writer.writerow([os.path.splitext(segment_filename)[0], speaker, subtitle.content])
|
267 |
+
|
268 |
+
# Clean up temporary files
|
269 |
+
cleanup(temp_path)
|
test2.py
ADDED
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
from helpers import *
|
4 |
+
from faster_whisper import WhisperModel
|
5 |
+
import whisperx
|
6 |
+
import torch
|
7 |
+
from pydub import AudioSegment
|
8 |
+
from nemo.collections.asr.models.msdd_models import NeuralDiarizer
|
9 |
+
import logging
|
10 |
+
import shutil
|
11 |
+
|
12 |
+
mtypes = {"cpu": "int8", "cuda": "float16"}
|
13 |
+
|
14 |
+
# Initialize parser
|
15 |
+
parser = argparse.ArgumentParser()
|
16 |
+
parser.add_argument(
|
17 |
+
"-a", "--audio", help="name of the target audio file", required=True
|
18 |
+
)
|
19 |
+
parser.add_argument(
|
20 |
+
"--no-stem",
|
21 |
+
action="store_false",
|
22 |
+
dest="stemming",
|
23 |
+
default=True,
|
24 |
+
help="Disables source separation. This helps with long files that don't contain a lot of music.",
|
25 |
+
)
|
26 |
+
parser.add_argument(
|
27 |
+
"--suppress_numerals",
|
28 |
+
action="store_true",
|
29 |
+
dest="suppress_numerals",
|
30 |
+
default=False,
|
31 |
+
help="Suppresses Numerical Digits. This helps the diarization accuracy but converts all digits into written text.",
|
32 |
+
)
|
33 |
+
parser.add_argument(
|
34 |
+
"--whisper-model",
|
35 |
+
dest="model_name",
|
36 |
+
default="medium.en",
|
37 |
+
help="name of the Whisper model to use",
|
38 |
+
)
|
39 |
+
parser.add_argument(
|
40 |
+
"--batch-size",
|
41 |
+
type=int,
|
42 |
+
dest="batch_size",
|
43 |
+
default=8,
|
44 |
+
help="Batch size for batched inference, reduce if you run out of memory, set to 0 for non-batched inference",
|
45 |
+
)
|
46 |
+
parser.add_argument(
|
47 |
+
"--language",
|
48 |
+
type=str,
|
49 |
+
default=None,
|
50 |
+
choices=whisper_langs,
|
51 |
+
help="Language spoken in the audio, specify None to perform language detection",
|
52 |
+
)
|
53 |
+
parser.add_argument(
|
54 |
+
"--device",
|
55 |
+
dest="device",
|
56 |
+
default="cuda" if torch.cuda.is_available() else "cpu",
|
57 |
+
help="if you have a GPU use 'cuda', otherwise 'cpu'",
|
58 |
+
)
|
59 |
+
args = parser.parse_args()
|
60 |
+
|
61 |
+
if args.stemming:
|
62 |
+
# Isolate vocals from the rest of the audio
|
63 |
+
return_code = os.system(
|
64 |
+
f'python3 -m demucs.separate -n htdemucs --two-stems=vocals "{args.audio}" -o "temp_outputs"'
|
65 |
+
)
|
66 |
+
if return_code != 0:
|
67 |
+
logging.warning(
|
68 |
+
"Source splitting failed, using original audio file. Use --no-stem argument to disable it."
|
69 |
+
)
|
70 |
+
vocal_target = args.audio
|
71 |
+
else:
|
72 |
+
vocal_target = os.path.join(
|
73 |
+
"temp_outputs",
|
74 |
+
"htdemucs",
|
75 |
+
os.path.splitext(os.path.basename(args.audio))[0],
|
76 |
+
"vocals.wav",
|
77 |
+
)
|
78 |
+
else:
|
79 |
+
vocal_target = args.audio
|
80 |
+
|
81 |
+
# Transcribe the audio file
|
82 |
+
if args.batch_size != 0:
|
83 |
+
from transcription_helpers import transcribe_batched
|
84 |
+
whisper_results, language = transcribe_batched(
|
85 |
+
vocal_target,
|
86 |
+
args.language,
|
87 |
+
args.batch_size,
|
88 |
+
args.model_name,
|
89 |
+
mtypes[args.device],
|
90 |
+
args.suppress_numerals,
|
91 |
+
args.device,
|
92 |
+
)
|
93 |
+
else:
|
94 |
+
from transcription_helpers import transcribe
|
95 |
+
whisper_results, language = transcribe(
|
96 |
+
vocal_target,
|
97 |
+
args.language,
|
98 |
+
args.model_name,
|
99 |
+
mtypes[args.device],
|
100 |
+
args.suppress_numerals,
|
101 |
+
args.device,
|
102 |
+
)
|
103 |
+
|
104 |
+
if language in wav2vec2_langs:
|
105 |
+
alignment_model, metadata = whisperx.load_align_model(
|
106 |
+
language_code=language, device=args.device
|
107 |
+
)
|
108 |
+
result_aligned = whisperx.align(
|
109 |
+
whisper_results, alignment_model, metadata, vocal_target, args.device
|
110 |
+
)
|
111 |
+
word_timestamps = filter_missing_timestamps(
|
112 |
+
result_aligned["word_segments"],
|
113 |
+
initial_timestamp=whisper_results[0].get("start"),
|
114 |
+
final_timestamp=whisper_results[-1].get("end"),
|
115 |
+
)
|
116 |
+
# clear gpu vram
|
117 |
+
del alignment_model
|
118 |
+
torch.cuda.empty_cache()
|
119 |
+
else:
|
120 |
+
assert (
|
121 |
+
args.batch_size == 0 # TODO: add a better check for word timestamps existence
|
122 |
+
), (
|
123 |
+
f"Unsupported language: {language}, use --batch_size to 0"
|
124 |
+
" to generate word timestamps using whisper directly and fix this error."
|
125 |
+
)
|
126 |
+
word_timestamps = []
|
127 |
+
for segment in whisper_results:
|
128 |
+
for word in segment["words"]:
|
129 |
+
word_timestamps.append({"word": word[2], "start": word[0], "end": word[1]})
|
130 |
+
|
131 |
+
# convert audio to mono for NeMo compatibility
|
132 |
+
sound = AudioSegment.from_file(vocal_target).set_channels(1)
|
133 |
+
ROOT = os.getcwd()
|
134 |
+
temp_path = os.path.join(ROOT, "temp_outputs")
|
135 |
+
os.makedirs(temp_path, exist_ok=True)
|
136 |
+
sound.export(os.path.join(temp_path, "mono_file.wav"), format="wav")
|
137 |
+
|
138 |
+
# Initialize NeMo MSDD diarization model
|
139 |
+
msdd_model = NeuralDiarizer(cfg=create_config(temp_path)).to(args.device)
|
140 |
+
msdd_model.diarize()
|
141 |
+
del msdd_model
|
142 |
+
torch.cuda.empty_cache()
|
143 |
+
|
144 |
+
# Reading timestamps <> Speaker Labels mapping
|
145 |
+
speaker_ts = []
|
146 |
+
with open(os.path.join(temp_path, "pred_rttms", "mono_file.rttm"), "r") as f:
|
147 |
+
lines = f.readlines()
|
148 |
+
for line in lines:
|
149 |
+
line_list = line.split(" ")
|
150 |
+
s = int(float(line_list[5]) * 1000)
|
151 |
+
e = s + int(float(line_list[8]) * 1000)
|
152 |
+
speaker_ts.append([s, e, int(line_list[11].split("_")[-1])])
|
153 |
+
|
154 |
+
wsm = get_words_speaker_mapping(word_timestamps, speaker_ts, "start")
|
155 |
+
wsm = get_realigned_ws_mapping_with_punctuation(wsm)
|
156 |
+
ssm = get_sentences_speaker_mapping(wsm, speaker_ts)
|
157 |
+
|
158 |
+
# Create the autodiarization directory structure
|
159 |
+
autodiarization_dir = "autodiarization"
|
160 |
+
os.makedirs(autodiarization_dir, exist_ok=True)
|
161 |
+
|
162 |
+
# Get the base name of the audio file
|
163 |
+
base_name = os.path.splitext(os.path.basename(args.audio))[0]
|
164 |
+
|
165 |
+
# Create a subdirectory for the current audio file
|
166 |
+
audio_dir = os.path.join(autodiarization_dir, base_name)
|
167 |
+
os.makedirs(audio_dir, exist_ok=True)
|
168 |
+
|
169 |
+
# Split the audio and create LJSpeech datasets for each speaker
|
170 |
+
for speaker_id in sorted(set(s[2] for s in speaker_ts)):
|
171 |
+
speaker_dir = os.path.join(audio_dir, f"speaker_{speaker_id}")
|
172 |
+
os.makedirs(speaker_dir, exist_ok=True)
|
173 |
+
|
174 |
+
speaker_segments = [s for s in ssm if s["speaker"] == speaker_id]
|
175 |
+
|
176 |
+
metadata = []
|
177 |
+
for i, segment in enumerate(speaker_segments, start=1):
|
178 |
+
start_time = segment["start"] / 1000
|
179 |
+
end_time = segment["end"] / 1000
|
180 |
+
transcript = " ".join(w["word"] for w in segment["words"])
|
181 |
+
|
182 |
+
# Split the audio segment
|
183 |
+
segment_audio = sound[start_time * 1000 : end_time * 1000]
|
184 |
+
segment_path = os.path.join(speaker_dir, f"speaker_{speaker_id}_{i:03d}.wav")
|
185 |
+
segment_audio.export(segment_path, format="wav")
|
186 |
+
|
187 |
+
metadata.append(f"speaker_{speaker_id}_{i:03d}|speaker_{speaker_id}|{transcript}")
|
188 |
+
|
189 |
+
# Write the metadata.csv file for the speaker
|
190 |
+
with open(os.path.join(speaker_dir, "metadata.csv"), "w", encoding="utf-8") as f:
|
191 |
+
f.write("\n".join(metadata))
|
192 |
+
|
193 |
+
# Write the full transcript and SRT files
|
194 |
+
with open(f"{os.path.splitext(args.audio)[0]}.txt", "w", encoding="utf-8") as f:
|
195 |
+
get_speaker_aware_transcript(ssm, f)
|
196 |
+
|
197 |
+
with open(f"{os.path.splitext(args.audio)[0]}.srt", "w", encoding="utf-8") as srt:
|
198 |
+
write_srt(ssm, srt)
|
199 |
+
|
200 |
+
# Clean up temporary files
|
201 |
+
cleanup(temp_path)
|
test3.py
ADDED
@@ -0,0 +1,214 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
from helpers import *
|
4 |
+
from faster_whisper import WhisperModel
|
5 |
+
import whisperx
|
6 |
+
import torch
|
7 |
+
from pydub import AudioSegment
|
8 |
+
from nemo.collections.asr.models.msdd_models import NeuralDiarizer
|
9 |
+
import logging
|
10 |
+
import shutil
|
11 |
+
import srt
|
12 |
+
|
13 |
+
mtypes = {"cpu": "int8", "cuda": "float16"}
|
14 |
+
|
15 |
+
# Initialize parser
|
16 |
+
parser = argparse.ArgumentParser()
|
17 |
+
parser.add_argument(
|
18 |
+
"-a", "--audio", help="name of the target audio file", required=True
|
19 |
+
)
|
20 |
+
parser.add_argument(
|
21 |
+
"--no-stem",
|
22 |
+
action="store_false",
|
23 |
+
dest="stemming",
|
24 |
+
default=True,
|
25 |
+
help="Disables source separation. This helps with long files that don't contain a lot of music.",
|
26 |
+
)
|
27 |
+
parser.add_argument(
|
28 |
+
"--suppress_numerals",
|
29 |
+
action="store_true",
|
30 |
+
dest="suppress_numerals",
|
31 |
+
default=False,
|
32 |
+
help="Suppresses Numerical Digits. This helps the diarization accuracy but converts all digits into written text.",
|
33 |
+
)
|
34 |
+
parser.add_argument(
|
35 |
+
"--whisper-model",
|
36 |
+
dest="model_name",
|
37 |
+
default="medium.en",
|
38 |
+
help="name of the Whisper model to use",
|
39 |
+
)
|
40 |
+
parser.add_argument(
|
41 |
+
"--batch-size",
|
42 |
+
type=int,
|
43 |
+
dest="batch_size",
|
44 |
+
default=8,
|
45 |
+
help="Batch size for batched inference, reduce if you run out of memory, set to 0 for non-batched inference",
|
46 |
+
)
|
47 |
+
parser.add_argument(
|
48 |
+
"--language",
|
49 |
+
type=str,
|
50 |
+
default=None,
|
51 |
+
choices=whisper_langs,
|
52 |
+
help="Language spoken in the audio, specify None to perform language detection",
|
53 |
+
)
|
54 |
+
parser.add_argument(
|
55 |
+
"--device",
|
56 |
+
dest="device",
|
57 |
+
default="cuda" if torch.cuda.is_available() else "cpu",
|
58 |
+
help="if you have a GPU use 'cuda', otherwise 'cpu'",
|
59 |
+
)
|
60 |
+
args = parser.parse_args()
|
61 |
+
|
62 |
+
if args.stemming:
|
63 |
+
# Isolate vocals from the rest of the audio
|
64 |
+
return_code = os.system(
|
65 |
+
f'python3 -m demucs.separate -n htdemucs --two-stems=vocals "{args.audio}" -o "temp_outputs"'
|
66 |
+
)
|
67 |
+
if return_code != 0:
|
68 |
+
logging.warning(
|
69 |
+
"Source splitting failed, using original audio file. Use --no-stem argument to disable it."
|
70 |
+
)
|
71 |
+
vocal_target = args.audio
|
72 |
+
else:
|
73 |
+
vocal_target = os.path.join(
|
74 |
+
"temp_outputs",
|
75 |
+
"htdemucs",
|
76 |
+
os.path.splitext(os.path.basename(args.audio))[0],
|
77 |
+
"vocals.wav",
|
78 |
+
)
|
79 |
+
else:
|
80 |
+
vocal_target = args.audio
|
81 |
+
|
82 |
+
# Transcribe the audio file
|
83 |
+
if args.batch_size != 0:
|
84 |
+
from transcription_helpers import transcribe_batched
|
85 |
+
whisper_results, language = transcribe_batched(
|
86 |
+
vocal_target,
|
87 |
+
args.language,
|
88 |
+
args.batch_size,
|
89 |
+
args.model_name,
|
90 |
+
mtypes[args.device],
|
91 |
+
args.suppress_numerals,
|
92 |
+
args.device,
|
93 |
+
)
|
94 |
+
else:
|
95 |
+
from transcription_helpers import transcribe
|
96 |
+
whisper_results, language = transcribe(
|
97 |
+
vocal_target,
|
98 |
+
args.language,
|
99 |
+
args.model_name,
|
100 |
+
mtypes[args.device],
|
101 |
+
args.suppress_numerals,
|
102 |
+
args.device,
|
103 |
+
)
|
104 |
+
|
105 |
+
if language in wav2vec2_langs:
|
106 |
+
alignment_model, metadata = whisperx.load_align_model(
|
107 |
+
language_code=language, device=args.device
|
108 |
+
)
|
109 |
+
result_aligned = whisperx.align(
|
110 |
+
whisper_results, alignment_model, metadata, vocal_target, args.device
|
111 |
+
)
|
112 |
+
word_timestamps = filter_missing_timestamps(
|
113 |
+
result_aligned["word_segments"],
|
114 |
+
initial_timestamp=whisper_results[0].get("start"),
|
115 |
+
final_timestamp=whisper_results[-1].get("end"),
|
116 |
+
)
|
117 |
+
# clear gpu vram
|
118 |
+
del alignment_model
|
119 |
+
torch.cuda.empty_cache()
|
120 |
+
else:
|
121 |
+
assert (
|
122 |
+
args.batch_size == 0 # TODO: add a better check for word timestamps existence
|
123 |
+
), (
|
124 |
+
f"Unsupported language: {language}, use --batch_size to 0"
|
125 |
+
" to generate word timestamps using whisper directly and fix this error."
|
126 |
+
)
|
127 |
+
word_timestamps = []
|
128 |
+
for segment in whisper_results:
|
129 |
+
for word in segment["words"]:
|
130 |
+
word_timestamps.append({"word": word[2], "start": word[0], "end": word[1]})
|
131 |
+
|
132 |
+
|
133 |
+
# convert audio to mono for NeMo compatibility
|
134 |
+
sound = AudioSegment.from_file(vocal_target).set_channels(1)
|
135 |
+
ROOT = os.getcwd()
|
136 |
+
temp_path = os.path.join(ROOT, "temp_outputs")
|
137 |
+
os.makedirs(temp_path, exist_ok=True)
|
138 |
+
sound.export(os.path.join(temp_path, "mono_file.wav"), format="wav")
|
139 |
+
|
140 |
+
# Initialize NeMo MSDD diarization model
|
141 |
+
msdd_model = NeuralDiarizer(cfg=create_config(temp_path)).to(args.device)
|
142 |
+
msdd_model.diarize()
|
143 |
+
del msdd_model
|
144 |
+
torch.cuda.empty_cache()
|
145 |
+
|
146 |
+
# Reading timestamps <> Speaker Labels mapping
|
147 |
+
speaker_ts = []
|
148 |
+
with open(os.path.join(temp_path, "pred_rttms", "mono_file.rttm"), "r") as f:
|
149 |
+
lines = f.readlines()
|
150 |
+
for line in lines:
|
151 |
+
line_list = line.split(" ")
|
152 |
+
s = int(float(line_list[5]) * 1000)
|
153 |
+
e = s + int(float(line_list[8]) * 1000)
|
154 |
+
speaker_ts.append([s, e, int(line_list[11].split("_")[-1])])
|
155 |
+
|
156 |
+
wsm = get_words_speaker_mapping(word_timestamps, speaker_ts, "start")
|
157 |
+
wsm = get_realigned_ws_mapping_with_punctuation(wsm)
|
158 |
+
ssm = get_sentences_speaker_mapping(wsm, speaker_ts)
|
159 |
+
|
160 |
+
# Create the autodiarization directory structure
|
161 |
+
autodiarization_dir = "autodiarization"
|
162 |
+
os.makedirs(autodiarization_dir, exist_ok=True)
|
163 |
+
|
164 |
+
# Get the base name of the audio file
|
165 |
+
base_name = os.path.splitext(os.path.basename(args.audio))[0]
|
166 |
+
|
167 |
+
# Create a subdirectory for the current audio file
|
168 |
+
audio_dir = os.path.join(autodiarization_dir, base_name)
|
169 |
+
os.makedirs(audio_dir, exist_ok=True)
|
170 |
+
|
171 |
+
# Create a dictionary to store speaker-specific metadata
|
172 |
+
speaker_metadata = {}
|
173 |
+
|
174 |
+
# Generate the SRT file
|
175 |
+
srt_file = f"{os.path.splitext(args.audio)[0]}.srt"
|
176 |
+
with open(srt_file, "w", encoding="utf-8") as f:
|
177 |
+
write_srt(ssm, f)
|
178 |
+
|
179 |
+
# Read the generated SRT file
|
180 |
+
with open(srt_file, "r", encoding="utf-8") as f:
|
181 |
+
srt_data = f.read()
|
182 |
+
|
183 |
+
# Parse the SRT data
|
184 |
+
srt_segments = list(srt.parse(srt_data))
|
185 |
+
|
186 |
+
# Process each segment in the SRT data
|
187 |
+
for segment in srt_segments:
|
188 |
+
start_time = segment.start.total_seconds() * 1000
|
189 |
+
end_time = segment.end.total_seconds() * 1000
|
190 |
+
speaker_name, transcript = segment.content.split(": ", 1)
|
191 |
+
|
192 |
+
# Extract the speaker ID from the speaker name
|
193 |
+
speaker_id = int(speaker_name.split(" ")[-1])
|
194 |
+
|
195 |
+
# Split the audio segment
|
196 |
+
segment_audio = sound[start_time:end_time]
|
197 |
+
segment_path = os.path.join(audio_dir, f"speaker_{speaker_id}", f"speaker_{speaker_id}_{segment.index:03d}.wav")
|
198 |
+
os.makedirs(os.path.dirname(segment_path), exist_ok=True)
|
199 |
+
segment_audio.export(segment_path, format="wav")
|
200 |
+
|
201 |
+
# Store the metadata for each speaker
|
202 |
+
if speaker_name not in speaker_metadata:
|
203 |
+
speaker_metadata[speaker_name] = []
|
204 |
+
speaker_metadata[speaker_name].append(f"speaker_{speaker_id}_{segment.index:03d}|{speaker_name}|{transcript}")
|
205 |
+
|
206 |
+
# Write the metadata.csv file for each speaker
|
207 |
+
for speaker_name, metadata in speaker_metadata.items():
|
208 |
+
speaker_id = int(speaker_name.split(" ")[-1])
|
209 |
+
speaker_dir = os.path.join(audio_dir, f"speaker_{speaker_id}")
|
210 |
+
with open(os.path.join(speaker_dir, "metadata.csv"), "w", encoding="utf-8") as f:
|
211 |
+
f.write("\n".join(metadata))
|
212 |
+
|
213 |
+
# Clean up temporary files
|
214 |
+
cleanup(temp_path)
|
transcription_helpers.py
ADDED
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
|
4 |
+
def transcribe(
|
5 |
+
audio_file: str,
|
6 |
+
language: str,
|
7 |
+
model_name: str,
|
8 |
+
compute_dtype: str,
|
9 |
+
suppress_numerals: bool,
|
10 |
+
device: str,
|
11 |
+
):
|
12 |
+
from faster_whisper import WhisperModel
|
13 |
+
from helpers import find_numeral_symbol_tokens, wav2vec2_langs
|
14 |
+
|
15 |
+
# Faster Whisper non-batched
|
16 |
+
# Run on GPU with FP16
|
17 |
+
whisper_model = WhisperModel(model_name, device=device, compute_type=compute_dtype)
|
18 |
+
|
19 |
+
# or run on GPU with INT8
|
20 |
+
# model = WhisperModel(model_size, device="cuda", compute_type="int8_float16")
|
21 |
+
# or run on CPU with INT8
|
22 |
+
# model = WhisperModel(model_size, device="cpu", compute_type="int8")
|
23 |
+
|
24 |
+
if suppress_numerals:
|
25 |
+
numeral_symbol_tokens = find_numeral_symbol_tokens(whisper_model.hf_tokenizer)
|
26 |
+
else:
|
27 |
+
numeral_symbol_tokens = None
|
28 |
+
|
29 |
+
if language is not None and language in wav2vec2_langs:
|
30 |
+
word_timestamps = False
|
31 |
+
else:
|
32 |
+
word_timestamps = True
|
33 |
+
|
34 |
+
segments, info = whisper_model.transcribe(
|
35 |
+
audio_file,
|
36 |
+
language=language,
|
37 |
+
beam_size=5,
|
38 |
+
word_timestamps=word_timestamps, # TODO: disable this if the language is supported by wav2vec2
|
39 |
+
suppress_tokens=numeral_symbol_tokens,
|
40 |
+
vad_filter=True,
|
41 |
+
)
|
42 |
+
whisper_results = []
|
43 |
+
for segment in segments:
|
44 |
+
whisper_results.append(segment._asdict())
|
45 |
+
# clear gpu vram
|
46 |
+
del whisper_model
|
47 |
+
torch.cuda.empty_cache()
|
48 |
+
return whisper_results, info.language
|
49 |
+
|
50 |
+
|
51 |
+
def transcribe_batched(
|
52 |
+
audio_file: str,
|
53 |
+
language: str,
|
54 |
+
batch_size: int,
|
55 |
+
model_name: str,
|
56 |
+
compute_dtype: str,
|
57 |
+
suppress_numerals: bool,
|
58 |
+
device: str,
|
59 |
+
):
|
60 |
+
import whisperx
|
61 |
+
|
62 |
+
# Faster Whisper batched
|
63 |
+
whisper_model = whisperx.load_model(
|
64 |
+
model_name,
|
65 |
+
device,
|
66 |
+
compute_type=compute_dtype,
|
67 |
+
asr_options={"suppress_numerals": suppress_numerals},
|
68 |
+
)
|
69 |
+
audio = whisperx.load_audio(audio_file)
|
70 |
+
result = whisper_model.transcribe(audio, language=language, batch_size=batch_size)
|
71 |
+
del whisper_model
|
72 |
+
torch.cuda.empty_cache()
|
73 |
+
return result["segments"], result["language"]
|
update_metadata.py
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
# Base directory for the LJ Speech-like structure
|
4 |
+
base_dir = "LJ_Speech_dataset"
|
5 |
+
|
6 |
+
# Recursively process each speaker subdirectory
|
7 |
+
for root, dirs, files in os.walk(base_dir):
|
8 |
+
for file in files:
|
9 |
+
if file == "metadata.csv":
|
10 |
+
metadata_file = os.path.join(root, file)
|
11 |
+
speaker_name = os.path.basename(root)
|
12 |
+
|
13 |
+
# Read the metadata file
|
14 |
+
with open(metadata_file, "r", encoding="utf-8") as f:
|
15 |
+
lines = f.readlines()
|
16 |
+
|
17 |
+
# Update the metadata lines
|
18 |
+
updated_lines = []
|
19 |
+
for line in lines:
|
20 |
+
parts = line.strip().split("|")
|
21 |
+
if len(parts) == 3:
|
22 |
+
parts[1] = speaker_name
|
23 |
+
updated_line = "|".join(parts)
|
24 |
+
updated_lines.append(updated_line)
|
25 |
+
|
26 |
+
# Write the updated metadata back to the file
|
27 |
+
with open(metadata_file, "w", encoding="utf-8") as f:
|
28 |
+
f.write("\n".join(updated_lines))
|
29 |
+
|
30 |
+
print(f"Updated metadata for {speaker_name}")
|
youtube_to_wav.py
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import unicode_literals
|
2 |
+
import yt_dlp
|
3 |
+
import ffmpeg
|
4 |
+
import sys
|
5 |
+
|
6 |
+
ydl_opts = {
|
7 |
+
'format': 'bestaudio/best',
|
8 |
+
# 'outtmpl': 'output.%(ext)s',
|
9 |
+
'postprocessors': [{
|
10 |
+
'key': 'FFmpegExtractAudio',
|
11 |
+
'preferredcodec': 'wav',
|
12 |
+
}],
|
13 |
+
}
|
14 |
+
def download_from_url(url):
|
15 |
+
ydl.download([url])
|
16 |
+
stream = ffmpeg.input('output.m4a')
|
17 |
+
stream = ffmpeg.output(stream, 'output.wav')
|
18 |
+
|
19 |
+
|
20 |
+
with yt_dlp.YoutubeDL(ydl_opts) as ydl:
|
21 |
+
args = sys.argv[1:]
|
22 |
+
if len(args) > 1:
|
23 |
+
print("Too many arguments.")
|
24 |
+
print("Usage: python youtubetowav.py <optional link>")
|
25 |
+
print("If a link is given it will automatically convert it to .wav. Otherwise a prompt will be shown")
|
26 |
+
exit()
|
27 |
+
if len(args) == 0:
|
28 |
+
url=input("Enter Youtube URL: ")
|
29 |
+
download_from_url(url)
|
30 |
+
else:
|
31 |
+
download_from_url(args[0])
|
32 |
+
|
33 |
+
|