Spaces:
Running
on
T4
Running
on
T4
fix_sampling_rate
#1
by
asahi417
- opened
- app.py +20 -77
- requirements.txt +1 -3
app.py
CHANGED
@@ -1,85 +1,37 @@
|
|
1 |
-
import os
|
2 |
-
import time
|
3 |
-
import tempfile
|
4 |
-
from math import floor
|
5 |
-
from typing import Optional, List, Dict, Any
|
6 |
-
|
7 |
import torch
|
|
|
8 |
import gradio as gr
|
9 |
import yt_dlp as youtube_dl
|
10 |
from transformers import pipeline
|
11 |
from transformers.pipelines.audio_utils import ffmpeg_read
|
12 |
|
|
|
|
|
13 |
|
14 |
-
|
15 |
-
|
16 |
-
BATCH_SIZE = 16
|
17 |
-
CHUNK_LENGTH_S = 15
|
18 |
FILE_LIMIT_MB = 1000
|
19 |
YT_LENGTH_LIMIT_S = 3600 # limit to 1 hour YouTube files
|
20 |
-
|
21 |
-
if torch.cuda.is_available()
|
22 |
-
torch_dtype = torch.bfloat16
|
23 |
-
device = "cuda:0"
|
24 |
-
model_kwargs = {'attn_implementation': 'sdpa'}
|
25 |
-
else:
|
26 |
-
torch_dtype = torch.float32
|
27 |
-
device = "cpu"
|
28 |
-
model_kwargs = {}
|
29 |
-
# define the pipeline
|
30 |
pipe = pipeline(
|
|
|
31 |
model=MODEL_NAME,
|
32 |
-
chunk_length_s=
|
33 |
-
batch_size=BATCH_SIZE,
|
34 |
-
torch_dtype=torch_dtype,
|
35 |
device=device,
|
36 |
-
model_kwargs=model_kwargs,
|
37 |
-
trust_remote_code=True
|
38 |
)
|
39 |
|
40 |
|
41 |
-
def
|
42 |
-
|
43 |
-
def _format_time(seconds: Optional[float]):
|
44 |
-
if seconds is None:
|
45 |
-
return "complete "
|
46 |
-
minutes = floor(seconds / 60)
|
47 |
-
hours = floor(seconds / 3600)
|
48 |
-
seconds = seconds - hours * 3600 - minutes * 60
|
49 |
-
m_seconds = floor(round(seconds - floor(seconds), 3) * 10 ** 3)
|
50 |
-
seconds = floor(seconds)
|
51 |
-
return f'{hours:02}:{minutes:02}:{seconds:02}.{m_seconds:03}'
|
52 |
-
|
53 |
-
return f"[{_format_time(start)}-> {_format_time(end)}]:"
|
54 |
-
|
55 |
-
|
56 |
-
def get_prediction(inputs, prompt: Optional[str]):
|
57 |
-
generate_kwargs = {"language": "japanese", "task": "transcribe"}
|
58 |
-
if prompt:
|
59 |
-
generate_kwargs['prompt_ids'] = pipe.tokenizer.get_prompt_ids(prompt, return_tensors='pt').to(device)
|
60 |
-
prediction = pipe(inputs, return_timestamps=True, generate_kwargs=generate_kwargs)
|
61 |
-
text = "".join([c['text'] for c in prediction['chunks']])
|
62 |
-
text_timestamped = "\n".join([
|
63 |
-
f"{format_time(*c['timestamp'])} {c['text']}" for c in prediction['chunks']
|
64 |
-
])
|
65 |
-
return text, text_timestamped
|
66 |
-
|
67 |
-
|
68 |
-
def transcribe(inputs: str, prompt):
|
69 |
if inputs is None:
|
70 |
raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
|
71 |
-
|
72 |
-
inputs = f.read()
|
73 |
-
inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
|
74 |
-
inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}
|
75 |
-
return get_prediction(inputs, prompt)
|
76 |
|
77 |
|
78 |
def _return_yt_html_embed(yt_url):
|
79 |
video_id = yt_url.split("?v=")[-1]
|
80 |
return f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe> </center>'
|
81 |
|
82 |
-
|
83 |
def download_yt_audio(yt_url, filename):
|
84 |
info_loader = youtube_dl.YoutubeDL()
|
85 |
try:
|
@@ -106,7 +58,7 @@ def download_yt_audio(yt_url, filename):
|
|
106 |
raise gr.Error(str(err))
|
107 |
|
108 |
|
109 |
-
def yt_transcribe(yt_url,
|
110 |
html_embed_str = _return_yt_html_embed(yt_url)
|
111 |
with tempfile.TemporaryDirectory() as tmpdirname:
|
112 |
filepath = os.path.join(tmpdirname, "video.mp4")
|
@@ -115,18 +67,15 @@ def yt_transcribe(yt_url, prompt):
|
|
115 |
inputs = f.read()
|
116 |
inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
|
117 |
inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}
|
118 |
-
text
|
119 |
-
return html_embed_str, text
|
120 |
|
121 |
|
122 |
demo = gr.Blocks()
|
123 |
mf_transcribe = gr.Interface(
|
124 |
fn=transcribe,
|
125 |
-
inputs=[
|
126 |
-
|
127 |
-
gr.inputs.Textbox(lines=1, placeholder="Prompt", optional=True),
|
128 |
-
],
|
129 |
-
outputs=["text", "text"],
|
130 |
layout="horizontal",
|
131 |
theme="huggingface",
|
132 |
title=f"Transcribe Audio with {os.path.basename(MODEL_NAME)}",
|
@@ -136,11 +85,8 @@ mf_transcribe = gr.Interface(
|
|
136 |
|
137 |
file_transcribe = gr.Interface(
|
138 |
fn=transcribe,
|
139 |
-
inputs=[
|
140 |
-
|
141 |
-
gr.inputs.Textbox(lines=1, placeholder="Prompt", optional=True),
|
142 |
-
],
|
143 |
-
outputs=["text", "text"],
|
144 |
layout="horizontal",
|
145 |
theme="huggingface",
|
146 |
title=f"Transcribe Audio with {os.path.basename(MODEL_NAME)}",
|
@@ -149,11 +95,8 @@ file_transcribe = gr.Interface(
|
|
149 |
)
|
150 |
yt_transcribe = gr.Interface(
|
151 |
fn=yt_transcribe,
|
152 |
-
inputs=[
|
153 |
-
|
154 |
-
gr.inputs.Textbox(lines=1, placeholder="Prompt", optional=True),
|
155 |
-
],
|
156 |
-
outputs=["html", "text", "text"],
|
157 |
layout="horizontal",
|
158 |
theme="huggingface",
|
159 |
title=f"Transcribe YouTube with {os.path.basename(MODEL_NAME)}",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import torch
|
2 |
+
|
3 |
import gradio as gr
|
4 |
import yt_dlp as youtube_dl
|
5 |
from transformers import pipeline
|
6 |
from transformers.pipelines.audio_utils import ffmpeg_read
|
7 |
|
8 |
+
import tempfile
|
9 |
+
import os
|
10 |
|
11 |
+
MODEL_NAME = "kotoba-tech/kotoba-whisper-v1.0"
|
12 |
+
BATCH_SIZE = 8
|
|
|
|
|
13 |
FILE_LIMIT_MB = 1000
|
14 |
YT_LENGTH_LIMIT_S = 3600 # limit to 1 hour YouTube files
|
15 |
+
|
16 |
+
device = 0 if torch.cuda.is_available() else "cpu"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
pipe = pipeline(
|
18 |
+
task="automatic-speech-recognition",
|
19 |
model=MODEL_NAME,
|
20 |
+
chunk_length_s=30,
|
|
|
|
|
21 |
device=device,
|
|
|
|
|
22 |
)
|
23 |
|
24 |
|
25 |
+
def transcribe(inputs):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
if inputs is None:
|
27 |
raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
|
28 |
+
return pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": "transcribe"}, return_timestamps=True)["text"]
|
|
|
|
|
|
|
|
|
29 |
|
30 |
|
31 |
def _return_yt_html_embed(yt_url):
|
32 |
video_id = yt_url.split("?v=")[-1]
|
33 |
return f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe> </center>'
|
34 |
|
|
|
35 |
def download_yt_audio(yt_url, filename):
|
36 |
info_loader = youtube_dl.YoutubeDL()
|
37 |
try:
|
|
|
58 |
raise gr.Error(str(err))
|
59 |
|
60 |
|
61 |
+
def yt_transcribe(yt_url, max_filesize=75.0):
|
62 |
html_embed_str = _return_yt_html_embed(yt_url)
|
63 |
with tempfile.TemporaryDirectory() as tmpdirname:
|
64 |
filepath = os.path.join(tmpdirname, "video.mp4")
|
|
|
67 |
inputs = f.read()
|
68 |
inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
|
69 |
inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}
|
70 |
+
text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": "transcribe"}, return_timestamps=True)["text"]
|
71 |
+
return html_embed_str, text
|
72 |
|
73 |
|
74 |
demo = gr.Blocks()
|
75 |
mf_transcribe = gr.Interface(
|
76 |
fn=transcribe,
|
77 |
+
inputs=[gr.inputs.Audio(source="microphone", type="filepath", optional=True)],
|
78 |
+
outputs="text",
|
|
|
|
|
|
|
79 |
layout="horizontal",
|
80 |
theme="huggingface",
|
81 |
title=f"Transcribe Audio with {os.path.basename(MODEL_NAME)}",
|
|
|
85 |
|
86 |
file_transcribe = gr.Interface(
|
87 |
fn=transcribe,
|
88 |
+
inputs=[gr.inputs.Audio(source="upload", type="filepath", optional=True, label="Audio file")],
|
89 |
+
outputs="text",
|
|
|
|
|
|
|
90 |
layout="horizontal",
|
91 |
theme="huggingface",
|
92 |
title=f"Transcribe Audio with {os.path.basename(MODEL_NAME)}",
|
|
|
95 |
)
|
96 |
yt_transcribe = gr.Interface(
|
97 |
fn=yt_transcribe,
|
98 |
+
inputs=[gr.inputs.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL")],
|
99 |
+
outputs=["html", "text"],
|
|
|
|
|
|
|
100 |
layout="horizontal",
|
101 |
theme="huggingface",
|
102 |
title=f"Transcribe YouTube with {os.path.basename(MODEL_NAME)}",
|
requirements.txt
CHANGED
@@ -1,5 +1,3 @@
|
|
1 |
git+https://github.com/huggingface/transformers
|
2 |
torch
|
3 |
-
yt-dlp
|
4 |
-
punctuators==0.0.5
|
5 |
-
stable-ts==2.16.0
|
|
|
1 |
git+https://github.com/huggingface/transformers
|
2 |
torch
|
3 |
+
yt-dlp
|
|
|
|