Dataset_STT

This dataset is designed for speech-to-text (STT) tasks and contains audio files along with their corresponding transcripts for the Uzbek language. The dataset is organized into two main directories: one for the audio files and another for the transcript files. Each of these directories is further divided by language (uz) and by the data split: train, test, and validation.

Dataset Structure

Dataset_STT/ ├── audio/ │ ├── uz/ │ │ ├── test/ │ │ │ └── test.tar │ │ ├── train/ │ │ │ └── train.tar │ │ └── validation/ │ │ └── validation.tar ├── transcript/ │ ├── uz/ │ │ ├── test/ │ │ │ └── test.tsv │ │ ├── train/ │ │ │ └── train.tsv │ │ └── validation/ │ │ └── validation.tsv

Files Description

Audio Files:
The audio files are stored as tar archives for each data split (train, test, and validation). Each tar archive contains the actual audio recordings (e.g., in MP3 format).
Transcript Files:
The transcript files are provided in TSV format with the following columns:
- id
- path (the filename of the audio file within the tar archive)
- sentence (the transcription)
- duration (audio duration in seconds)
- age
- gender
- accents
- locale
Custom Loader (dataset_stt.py):
This script extracts audio files from the tar archives and pairs them with their metadata from the transcript TSV files. It returns the audio data using the datasets.Audio feature (with a sampling rate of 16000 Hz), which enables interactive playback in the Hugging Face dataset viewer.

How to Load the Dataset

You can load the dataset using the Hugging Face datasets library. For example:

from datasets import load_dataset

data_files = {
    "train": {"audio": "audio/uz/train/train.tar", "transcript": "transcript/uz/train/train.tsv"},
    "test": {"audio": "audio/uz/test/test.tar", "transcript": "transcript/uz/test/test.tsv"},
    "validation": {"audio": "audio/uz/validation/validation.tar", "transcript": "transcript/uz/validation/validation.tsv"}
}

dataset = load_dataset("Elyordev/Dataset_STT", data_files=data_files)
print(dataset)