Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,129 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from nemo.collections.asr.models import EncDecMultiTaskModel
|
2 |
+
import gradio as gr
|
3 |
+
import torch
|
4 |
+
import json
|
5 |
+
import numpy as np
|
6 |
+
import soundfile as sf
|
7 |
+
import tempfile
|
8 |
+
from transformers import VitsTokenizer, VitsModel, set_seed
|
9 |
+
|
10 |
+
|
11 |
+
|
12 |
+
#just to import this piece of shit above me, one needs:
|
13 |
+
|
14 |
+
#gradio transformers
|
15 |
+
#nemo
|
16 |
+
#hydra
|
17 |
+
#librosa
|
18 |
+
#sentencepiece
|
19 |
+
#
|
20 |
+
#
|
21 |
+
|
22 |
+
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
|
27 |
+
|
28 |
+
# load model
|
29 |
+
canary_model = EncDecMultiTaskModel.from_pretrained('nvidia/canary-1b')
|
30 |
+
|
31 |
+
# update decode params
|
32 |
+
decode_cfg = canary_model.cfg.decoding
|
33 |
+
decode_cfg.beam.beam_size = 1
|
34 |
+
canary_model.change_decoding_strategy(decode_cfg)
|
35 |
+
|
36 |
+
|
37 |
+
import torch
|
38 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
|
39 |
+
|
40 |
+
|
41 |
+
#install accelerate
|
42 |
+
|
43 |
+
torch.random.manual_seed(0)
|
44 |
+
|
45 |
+
model = AutoModelForCausalLM.from_pretrained(
|
46 |
+
"microsoft/Phi-3-mini-128k-instruct",
|
47 |
+
device_map="cpu",
|
48 |
+
torch_dtype="auto",
|
49 |
+
trust_remote_code=True,
|
50 |
+
)
|
51 |
+
tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-128k-instruct")
|
52 |
+
|
53 |
+
messages = []
|
54 |
+
|
55 |
+
pipe = pipeline(
|
56 |
+
"text-generation",
|
57 |
+
model=model,
|
58 |
+
tokenizer=tokenizer,
|
59 |
+
)
|
60 |
+
|
61 |
+
generation_args = {
|
62 |
+
"max_new_tokens": 500,
|
63 |
+
"return_full_text": False,
|
64 |
+
"temperature": 0.0,
|
65 |
+
"do_sample": False,
|
66 |
+
}
|
67 |
+
|
68 |
+
|
69 |
+
tokenizer_vits = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")
|
70 |
+
model_vits = VitsModel.from_pretrained("facebook/mms-tts-eng")
|
71 |
+
|
72 |
+
# Define the function to transcribe audio
|
73 |
+
def transcribe_audio(audio):
|
74 |
+
audio_list, sample_rate = sf.read(audio)
|
75 |
+
|
76 |
+
if audio_list.ndim > 1:
|
77 |
+
audio_list = np.mean(audio_list,axis=1)
|
78 |
+
|
79 |
+
# Create a temporary file to save the audio data
|
80 |
+
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_audio_file:
|
81 |
+
temp_audio_path = temp_audio_file.name
|
82 |
+
|
83 |
+
# Save the audio data to the temporary file
|
84 |
+
sf.write(temp_audio_path, audio_list, sample_rate)
|
85 |
+
|
86 |
+
# Transcribe audio using the canary model
|
87 |
+
predicted_text = canary_model.transcribe(paths2audio_files=[temp_audio_path], batch_size=16)
|
88 |
+
|
89 |
+
# Remove the temporary file
|
90 |
+
|
91 |
+
# Return the transcription
|
92 |
+
messages = [{"role": "user", "content": predicted_text[0]}]
|
93 |
+
|
94 |
+
output_text =pipe(messages, **generation_args)
|
95 |
+
|
96 |
+
inputs_vits = tokenizer_vits(text=output_text[0]["generated_text"], return_tensors="pt")
|
97 |
+
|
98 |
+
set_seed(555) # make deterministic
|
99 |
+
|
100 |
+
with torch.no_grad():
|
101 |
+
outputs_vits = model_vits(**inputs_vits)
|
102 |
+
|
103 |
+
waveform = outputs_vits.waveform[0]
|
104 |
+
|
105 |
+
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_audio_file_2:
|
106 |
+
temp_audio_path_2 = temp_audio_file_2.name
|
107 |
+
|
108 |
+
# Save the audio data to the temporary file
|
109 |
+
sf.write(temp_audio_path_2, waveform.numpy(), model_vits.config.sampling_rate)
|
110 |
+
|
111 |
+
return temp_audio_path_2
|
112 |
+
|
113 |
+
|
114 |
+
|
115 |
+
|
116 |
+
# Create the Gradio interface
|
117 |
+
import gradio as gr
|
118 |
+
|
119 |
+
|
120 |
+
|
121 |
+
|
122 |
+
|
123 |
+
#gradio replaced .input and .output with .components
|
124 |
+
audio_input = gr.components.Audio(sources=["upload","microphone"], type="filepath", label="Record Audio")
|
125 |
+
audio_output = gr.components.Audio(label="Audio Output")
|
126 |
+
interface = gr.Interface(fn=transcribe_audio, inputs=audio_input, outputs=audio_output)
|
127 |
+
|
128 |
+
# Launch the interface
|
129 |
+
interface.launch()
|