Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,102 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import numpy as np
|
3 |
+
import os, time, librosa, torch
|
4 |
+
from pyannote.audio import Pipeline
|
5 |
+
from transformers import pipeline
|
6 |
+
from utils import second_to_timecode, download_from_youtube
|
7 |
+
|
8 |
+
MODEL_NAME = 'openai/whisper-medium'
|
9 |
+
lang = 'en'
|
10 |
+
|
11 |
+
chunk_length_s = 9
|
12 |
+
vad_activation_min_duration = 9 # sec
|
13 |
+
device = 0 if torch.cuda.is_available() else "cpu"
|
14 |
+
SAMPLE_RATE = 16_000
|
15 |
+
|
16 |
+
######## LOAD MODELS FROM HUB ########
|
17 |
+
dia_model = Pipeline.from_pretrained("pyannote/speaker-diarization", use_auth_token=os.environ['TOKEN'])
|
18 |
+
vad_model = Pipeline.from_pretrained("pyannote/voice-activity-detection", use_auth_token=os.environ['TOKEN'])
|
19 |
+
pipe = pipeline(task="automatic-speech-recognition", model=MODEL_NAME, chunk_length_s=chunk_length_s, device=device)
|
20 |
+
pipe.model.config.forced_decoder_ids = pipe.tokenizer.get_decoder_prompt_ids(language=lang, task="transcribe")
|
21 |
+
|
22 |
+
print("----------> Loaded models <-----------")
|
23 |
+
|
24 |
+
def generator(youtube_link, microphone, file_upload, num_speakers, max_duration, history):
|
25 |
+
|
26 |
+
if int(youtube_link != '') + int(microphone is not None) + int(file_upload is not None) != 1:
|
27 |
+
raise Exception(f"Only one of the source should be given youtube_link={youtube_link}, microphone={microphone}, file_upload={file_upload}")
|
28 |
+
|
29 |
+
history = history or ""
|
30 |
+
|
31 |
+
if microphone:
|
32 |
+
path = microphone
|
33 |
+
elif file_upload:
|
34 |
+
path = file_upload
|
35 |
+
elif youtube_link:
|
36 |
+
path = download_from_youtube(youtube_link)
|
37 |
+
|
38 |
+
waveform, sampling_rate = librosa.load(path, sr=SAMPLE_RATE, mono=True, duration=max_duration)
|
39 |
+
|
40 |
+
print(waveform.shape, sampling_rate)
|
41 |
+
waveform_tensor = torch.unsqueeze(torch.tensor(waveform), 0).to(device)
|
42 |
+
|
43 |
+
dia_result = dia_model({
|
44 |
+
"waveform": waveform_tensor,
|
45 |
+
"sample_rate": sampling_rate,
|
46 |
+
}, num_speakers=num_speakers)
|
47 |
+
|
48 |
+
for speech_turn, track, speaker in dia_result.itertracks(yield_label=True):
|
49 |
+
print(f"{speech_turn.start:4.1f} {speech_turn.end:4.1f} {speaker}")
|
50 |
+
_start = int(sampling_rate * speech_turn.start)
|
51 |
+
_end = int(sampling_rate * speech_turn.end)
|
52 |
+
data = waveform[_start: _end]
|
53 |
+
|
54 |
+
if speech_turn.end - speech_turn.start > vad_activation_min_duration:
|
55 |
+
print(f'audio duration {speech_turn.end - speech_turn.start} sec ----> activating VAD')
|
56 |
+
vad_output = vad_model({
|
57 |
+
'waveform': waveform_tensor[:, _start:_end],
|
58 |
+
'sample_rate': sampling_rate})
|
59 |
+
for vad_turn in vad_output.get_timeline().support():
|
60 |
+
vad_start = _start + int(sampling_rate * vad_turn.start)
|
61 |
+
vad_end = _start + int(sampling_rate * vad_turn.end)
|
62 |
+
prediction = pipe(waveform[vad_start: vad_end])['text']
|
63 |
+
history += f"{second_to_timecode(speech_turn.start + vad_turn.start)},{second_to_timecode(speech_turn.start + vad_turn.end)}\n" + \
|
64 |
+
f"{prediction}\n\n"
|
65 |
+
# f">> {speaker}: {prediction}\n\n"
|
66 |
+
yield history, history, None
|
67 |
+
|
68 |
+
else:
|
69 |
+
prediction = pipe(data)['text']
|
70 |
+
history += f"{second_to_timecode(speech_turn.start)},{second_to_timecode(speech_turn.end)}\n" + \
|
71 |
+
f"{prediction}\n\n"
|
72 |
+
# f">> {speaker}: {prediction}\n\n"
|
73 |
+
|
74 |
+
yield history, history, None
|
75 |
+
|
76 |
+
# https://support.google.com/youtube/answer/2734698?hl=en#zippy=%2Cbasic-file-formats%2Csubrip-srt-example%2Csubviewer-sbv-example
|
77 |
+
file_name = 'transcript.sbv'
|
78 |
+
with open(file_name, 'w') as fp:
|
79 |
+
fp.write(history)
|
80 |
+
|
81 |
+
yield history, history, file_name
|
82 |
+
|
83 |
+
demo = gr.Interface(
|
84 |
+
generator,
|
85 |
+
inputs=[
|
86 |
+
gr.inputs.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL", optional=True),
|
87 |
+
gr.inputs.Audio(source="microphone", type="filepath", optional=True),
|
88 |
+
gr.inputs.Audio(source="upload", type="filepath", optional=True),
|
89 |
+
gr.Number(value=1, label="Number of Speakers"),
|
90 |
+
gr.Number(value=120, label="Maximum Duration (Seconds)"),
|
91 |
+
'state',
|
92 |
+
],
|
93 |
+
outputs=['text', 'state', 'file'],
|
94 |
+
layout="horizontal",
|
95 |
+
theme="huggingface",
|
96 |
+
allow_flagging="never",
|
97 |
+
)
|
98 |
+
|
99 |
+
# define queue - required for generators
|
100 |
+
demo.queue()
|
101 |
+
|
102 |
+
demo.launch()
|