Spaces:
Running
Running
jhj0517
commited on
Commit
•
a63d493
1
Parent(s):
6726c6a
refactoring
Browse files- modules/whisper_Inference.py +148 -97
modules/whisper_Inference.py
CHANGED
@@ -1,7 +1,11 @@
|
|
1 |
import whisper
|
2 |
import gradio as gr
|
|
|
3 |
import os
|
|
|
|
|
4 |
from datetime import datetime
|
|
|
5 |
|
6 |
from .base_interface import BaseInterface
|
7 |
from modules.subtitle_manager import get_srt, get_vtt, write_file, safe_filename
|
@@ -48,61 +52,45 @@ class WhisperInference(BaseInterface):
|
|
48 |
Indicator to show progress directly in gradio.
|
49 |
I use a forked version of whisper for this. To see more info : https://github.com/jhj0517/jhj0517-whisper/tree/add-progress-callback
|
50 |
"""
|
51 |
-
def progress_callback(progress_value):
|
52 |
-
progress(progress_value, desc="Transcribing..")
|
53 |
|
54 |
try:
|
55 |
if model_size != self.current_model_size or self.model is None:
|
56 |
-
|
57 |
-
self.current_model_size = model_size
|
58 |
-
self.model = whisper.load_model(name=model_size, download_root=os.path.join("models", "Whisper"))
|
59 |
-
|
60 |
-
if lang == "Automatic Detection":
|
61 |
-
lang = None
|
62 |
-
|
63 |
-
progress(0, desc="Loading Audio..")
|
64 |
|
65 |
files_info = {}
|
66 |
for fileobj in fileobjs:
|
67 |
-
|
68 |
audio = whisper.load_audio(fileobj.name)
|
69 |
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
else:
|
75 |
-
result = self.model.transcribe(audio=audio, language=lang, verbose=False,
|
76 |
-
progress_callback=progress_callback)
|
77 |
-
|
78 |
progress(1, desc="Completed!")
|
79 |
|
80 |
file_name, file_ext = os.path.splitext(os.path.basename(fileobj.orig_name))
|
81 |
file_name = safe_filename(file_name)
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
|
|
87 |
|
88 |
-
|
89 |
-
subtitle = get_srt(result["segments"])
|
90 |
-
write_file(subtitle, f"{output_path}.srt")
|
91 |
-
elif subformat == "WebVTT":
|
92 |
-
subtitle = get_vtt(result["segments"])
|
93 |
-
write_file(subtitle, f"{output_path}.vtt")
|
94 |
-
|
95 |
-
files_info[file_name] = subtitle
|
96 |
|
97 |
total_result = ''
|
98 |
-
|
|
|
99 |
total_result += '------------------------------------\n'
|
100 |
total_result += f'{file_name}\n\n'
|
101 |
-
total_result += f'
|
|
|
102 |
|
103 |
-
return f"Done! Subtitle is in the outputs folder.\n\n{total_result}"
|
104 |
except Exception as e:
|
105 |
-
|
|
|
106 |
finally:
|
107 |
self.release_cuda_memory()
|
108 |
self.remove_input_files([fileobj.name for fileobj in fileobjs])
|
@@ -137,49 +125,32 @@ class WhisperInference(BaseInterface):
|
|
137 |
Indicator to show progress directly in gradio.
|
138 |
I use a forked version of whisper for this. To see more info : https://github.com/jhj0517/jhj0517-whisper/tree/add-progress-callback
|
139 |
"""
|
140 |
-
def progress_callback(progress_value):
|
141 |
-
progress(progress_value, desc="Transcribing..")
|
142 |
-
|
143 |
try:
|
144 |
if model_size != self.current_model_size or self.model is None:
|
145 |
-
|
146 |
-
self.current_model_size = model_size
|
147 |
-
self.model = whisper.load_model(name=model_size, download_root=os.path.join("models", "Whisper"))
|
148 |
-
|
149 |
-
if lang == "Automatic Detection":
|
150 |
-
lang = None
|
151 |
|
152 |
progress(0, desc="Loading Audio from Youtube..")
|
153 |
yt = get_ytdata(youtubelink)
|
154 |
audio = whisper.load_audio(get_ytaudio(yt))
|
155 |
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
else:
|
161 |
-
result = self.model.transcribe(audio=audio, language=lang, verbose=False,
|
162 |
-
progress_callback=progress_callback)
|
163 |
-
|
164 |
progress(1, desc="Completed!")
|
165 |
|
166 |
file_name = safe_filename(yt.title)
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
write_file(subtitle, f"{output_path}.srt")
|
176 |
-
elif subformat == "WebVTT":
|
177 |
-
subtitle = get_vtt(result["segments"])
|
178 |
-
write_file(subtitle, f"{output_path}.vtt")
|
179 |
-
|
180 |
-
return f"Done! Subtitle file is in the outputs folder.\n\n{subtitle}"
|
181 |
except Exception as e:
|
182 |
-
|
|
|
183 |
finally:
|
184 |
yt = get_ytdata(youtubelink)
|
185 |
file_path = get_ytaudio(yt)
|
@@ -213,43 +184,123 @@ class WhisperInference(BaseInterface):
|
|
213 |
Indicator to show progress directly in gradio.
|
214 |
I use a forked version of whisper for this. To see more info : https://github.com/jhj0517/jhj0517-whisper/tree/add-progress-callback
|
215 |
"""
|
216 |
-
def progress_callback(progress_value):
|
217 |
-
progress(progress_value, desc="Transcribing..")
|
218 |
|
219 |
try:
|
220 |
if model_size != self.current_model_size or self.model is None:
|
221 |
-
|
222 |
-
self.current_model_size = model_size
|
223 |
-
self.model = whisper.load_model(name=model_size, download_root=os.path.join("models", "Whisper"))
|
224 |
-
|
225 |
-
if lang == "Automatic Detection":
|
226 |
-
lang = None
|
227 |
-
|
228 |
-
progress(0, desc="Loading Audio..")
|
229 |
-
|
230 |
-
translatable_model = ["large", "large-v1", "large-v2"]
|
231 |
-
if istranslate and self.current_model_size in translatable_model:
|
232 |
-
result = self.model.transcribe(audio=micaudio, language=lang, verbose=False, task="translate",
|
233 |
-
progress_callback=progress_callback)
|
234 |
-
else:
|
235 |
-
result = self.model.transcribe(audio=micaudio, language=lang, verbose=False,
|
236 |
-
progress_callback=progress_callback)
|
237 |
|
|
|
|
|
|
|
|
|
238 |
progress(1, desc="Completed!")
|
239 |
|
240 |
-
|
241 |
-
|
|
|
|
|
|
|
|
|
242 |
|
243 |
-
|
244 |
-
subtitle = get_srt(result["segments"])
|
245 |
-
write_file(subtitle, f"{output_path}.srt")
|
246 |
-
elif subformat == "WebVTT":
|
247 |
-
subtitle = get_vtt(result["segments"])
|
248 |
-
write_file(subtitle, f"{output_path}.vtt")
|
249 |
-
|
250 |
-
return f"Done! Subtitle file is in the outputs folder.\n\n{subtitle}"
|
251 |
except Exception as e:
|
252 |
-
|
|
|
253 |
finally:
|
254 |
self.release_cuda_memory()
|
255 |
self.remove_input_files([micaudio])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import whisper
|
2 |
import gradio as gr
|
3 |
+
import time
|
4 |
import os
|
5 |
+
from typing import BinaryIO, Union, Tuple
|
6 |
+
import numpy as np
|
7 |
from datetime import datetime
|
8 |
+
import torch
|
9 |
|
10 |
from .base_interface import BaseInterface
|
11 |
from modules.subtitle_manager import get_srt, get_vtt, write_file, safe_filename
|
|
|
52 |
Indicator to show progress directly in gradio.
|
53 |
I use a forked version of whisper for this. To see more info : https://github.com/jhj0517/jhj0517-whisper/tree/add-progress-callback
|
54 |
"""
|
|
|
|
|
55 |
|
56 |
try:
|
57 |
if model_size != self.current_model_size or self.model is None:
|
58 |
+
self.initialize_model(model_size=model_size, progress=progress)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
|
60 |
files_info = {}
|
61 |
for fileobj in fileobjs:
|
62 |
+
progress(0, desc="Loading Audio..")
|
63 |
audio = whisper.load_audio(fileobj.name)
|
64 |
|
65 |
+
result, elapsed_time = self.transcribe(audio=audio,
|
66 |
+
lang=lang,
|
67 |
+
istranslate=istranslate,
|
68 |
+
progress=progress)
|
|
|
|
|
|
|
|
|
69 |
progress(1, desc="Completed!")
|
70 |
|
71 |
file_name, file_ext = os.path.splitext(os.path.basename(fileobj.orig_name))
|
72 |
file_name = safe_filename(file_name)
|
73 |
+
subtitle = self.generate_and_write_subtitle(
|
74 |
+
file_name=file_name,
|
75 |
+
transcribed_segments=result,
|
76 |
+
add_timestamp=add_timestamp,
|
77 |
+
subformat=subformat
|
78 |
+
)
|
79 |
|
80 |
+
files_info[file_name] = {"subtitle": subtitle, "elapsed_time": elapsed_time}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
total_result = ''
|
83 |
+
total_time = 0
|
84 |
+
for file_name, info in files_info.items():
|
85 |
total_result += '------------------------------------\n'
|
86 |
total_result += f'{file_name}\n\n'
|
87 |
+
total_result += f"{info['subtitle']}"
|
88 |
+
total_time += info["elapsed_time"]
|
89 |
|
90 |
+
return f"Done in {self.format_time(total_time)}! Subtitle is in the outputs folder.\n\n{total_result}"
|
91 |
except Exception as e:
|
92 |
+
print(f"Error transcribing file: {str(e)}")
|
93 |
+
return f"Error transcribing file: {str(e)}"
|
94 |
finally:
|
95 |
self.release_cuda_memory()
|
96 |
self.remove_input_files([fileobj.name for fileobj in fileobjs])
|
|
|
125 |
Indicator to show progress directly in gradio.
|
126 |
I use a forked version of whisper for this. To see more info : https://github.com/jhj0517/jhj0517-whisper/tree/add-progress-callback
|
127 |
"""
|
|
|
|
|
|
|
128 |
try:
|
129 |
if model_size != self.current_model_size or self.model is None:
|
130 |
+
self.initialize_model(model_size=model_size, progress=progress)
|
|
|
|
|
|
|
|
|
|
|
131 |
|
132 |
progress(0, desc="Loading Audio from Youtube..")
|
133 |
yt = get_ytdata(youtubelink)
|
134 |
audio = whisper.load_audio(get_ytaudio(yt))
|
135 |
|
136 |
+
result, elapsed_time = self.transcribe(audio=audio,
|
137 |
+
lang=lang,
|
138 |
+
istranslate=istranslate,
|
139 |
+
progress=progress)
|
|
|
|
|
|
|
|
|
140 |
progress(1, desc="Completed!")
|
141 |
|
142 |
file_name = safe_filename(yt.title)
|
143 |
+
subtitle = self.generate_and_write_subtitle(
|
144 |
+
file_name=file_name,
|
145 |
+
transcribed_segments=result,
|
146 |
+
add_timestamp=add_timestamp,
|
147 |
+
subformat=subformat
|
148 |
+
)
|
149 |
+
|
150 |
+
return f"Done in {self.format_time(elapsed_time)}! Subtitle file is in the outputs folder.\n\n{subtitle}"
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
except Exception as e:
|
152 |
+
print(f"Error transcribing youtube video: {str(e)}")
|
153 |
+
return f"Error transcribing youtube video: {str(e)}"
|
154 |
finally:
|
155 |
yt = get_ytdata(youtubelink)
|
156 |
file_path = get_ytaudio(yt)
|
|
|
184 |
Indicator to show progress directly in gradio.
|
185 |
I use a forked version of whisper for this. To see more info : https://github.com/jhj0517/jhj0517-whisper/tree/add-progress-callback
|
186 |
"""
|
|
|
|
|
187 |
|
188 |
try:
|
189 |
if model_size != self.current_model_size or self.model is None:
|
190 |
+
self.initialize_model(model_size=model_size, progress=progress)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
191 |
|
192 |
+
result, elapsed_time = self.transcribe(audio=micaudio,
|
193 |
+
lang=lang,
|
194 |
+
istranslate=istranslate,
|
195 |
+
progress=progress)
|
196 |
progress(1, desc="Completed!")
|
197 |
|
198 |
+
subtitle = self.generate_and_write_subtitle(
|
199 |
+
file_name="Mic",
|
200 |
+
transcribed_segments=result,
|
201 |
+
add_timestamp=True,
|
202 |
+
subformat=subformat
|
203 |
+
)
|
204 |
|
205 |
+
return f"Done in {self.format_time(elapsed_time)}! Subtitle file is in the outputs folder.\n\n{subtitle}"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
206 |
except Exception as e:
|
207 |
+
print(f"Error transcribing mic: {str(e)}")
|
208 |
+
return f"Error transcribing mic: {str(e)}"
|
209 |
finally:
|
210 |
self.release_cuda_memory()
|
211 |
self.remove_input_files([micaudio])
|
212 |
+
|
213 |
+
def transcribe(self,
|
214 |
+
audio: Union[str, np.ndarray, torch.Tensor],
|
215 |
+
lang: str,
|
216 |
+
istranslate: bool,
|
217 |
+
progress: gr.Progress
|
218 |
+
) -> Tuple[list[dict], float]:
|
219 |
+
"""
|
220 |
+
transcribe method for OpenAI's Whisper implementation.
|
221 |
+
|
222 |
+
Parameters
|
223 |
+
----------
|
224 |
+
audio: Union[str, BinaryIO, torch.Tensor]
|
225 |
+
Audio path or file binary or Audio numpy array
|
226 |
+
lang: str
|
227 |
+
Source language of the file to transcribe from gr.Dropdown()
|
228 |
+
istranslate: bool
|
229 |
+
Boolean value from gr.Checkbox() that determines whether to translate to English.
|
230 |
+
It's Whisper's feature to translate speech from another language directly into English end-to-end.
|
231 |
+
progress: gr.Progress
|
232 |
+
Indicator to show progress directly in gradio.
|
233 |
+
|
234 |
+
Returns
|
235 |
+
----------
|
236 |
+
segments_result: list[dict]
|
237 |
+
list of dicts that includes start, end timestamps and transcribed text
|
238 |
+
elapsed_time: float
|
239 |
+
elapsed time for transcription
|
240 |
+
"""
|
241 |
+
start_time = time.time()
|
242 |
+
|
243 |
+
def progress_callback(progress_value):
|
244 |
+
progress(progress_value, desc="Transcribing..")
|
245 |
+
|
246 |
+
if lang == "Automatic Detection":
|
247 |
+
lang = None
|
248 |
+
|
249 |
+
translatable_model = ["large", "large-v1", "large-v2"]
|
250 |
+
segments_result = self.model.transcribe(audio=audio,
|
251 |
+
language=lang,
|
252 |
+
verbose=False,
|
253 |
+
task="translate" if istranslate and self.current_model_size in translatable_model else "transcribe",
|
254 |
+
progress_callback=progress_callback)["segments"]
|
255 |
+
elapsed_time = time.time() - start_time
|
256 |
+
|
257 |
+
return segments_result, elapsed_time
|
258 |
+
|
259 |
+
def initialize_model(self,
|
260 |
+
model_size: str,
|
261 |
+
progress: gr.Progress
|
262 |
+
):
|
263 |
+
"""
|
264 |
+
Initialize model if it doesn't match with current model size
|
265 |
+
"""
|
266 |
+
progress(0, desc="Initializing Model..")
|
267 |
+
self.current_model_size = model_size
|
268 |
+
self.model = whisper.load_model(name=model_size, download_root=os.path.join("models", "Whisper"))
|
269 |
+
|
270 |
+
@staticmethod
|
271 |
+
def generate_and_write_subtitle(file_name: str,
|
272 |
+
transcribed_segments: list,
|
273 |
+
add_timestamp: bool,
|
274 |
+
subformat: str,
|
275 |
+
) -> str:
|
276 |
+
"""
|
277 |
+
This method writes subtitle file and returns str to gr.Textbox
|
278 |
+
"""
|
279 |
+
timestamp = datetime.now().strftime("%m%d%H%M%S")
|
280 |
+
if add_timestamp:
|
281 |
+
output_path = os.path.join("outputs", f"{file_name}-{timestamp}")
|
282 |
+
else:
|
283 |
+
output_path = os.path.join("outputs", f"{file_name}")
|
284 |
+
|
285 |
+
if subformat == "SRT":
|
286 |
+
subtitle = get_srt(transcribed_segments)
|
287 |
+
write_file(subtitle, f"{output_path}.srt")
|
288 |
+
elif subformat == "WebVTT":
|
289 |
+
subtitle = get_vtt(transcribed_segments)
|
290 |
+
write_file(subtitle, f"{output_path}.vtt")
|
291 |
+
return subtitle
|
292 |
+
|
293 |
+
@staticmethod
|
294 |
+
def format_time(elapsed_time: float) -> str:
|
295 |
+
hours, rem = divmod(elapsed_time, 3600)
|
296 |
+
minutes, seconds = divmod(rem, 60)
|
297 |
+
|
298 |
+
time_str = ""
|
299 |
+
if hours:
|
300 |
+
time_str += f"{hours} hours "
|
301 |
+
if minutes:
|
302 |
+
time_str += f"{minutes} minutes "
|
303 |
+
seconds = round(seconds)
|
304 |
+
time_str += f"{seconds} seconds"
|
305 |
+
|
306 |
+
return time_str.strip()
|