File size: 38,795 Bytes
cb9ee50
b84f4f2
31f7bdb
33a2c1e
95261ed
8d120bf
05a2178
3fadc6e
8d120bf
3fadc6e
cb9ee50
fdb8dbd
20f75ae
 
295de00
74b7d77
a1da02d
295de00
 
 
adca588
c0e541b
74b7d77
 
cb9ee50
31f7bdb
95261ed
8d120bf
 
 
 
05a2178
 
883c794
43189ac
95261ed
295de00
 
7ce6041
1acaa19
533d92e
6a308c6
 
 
20f75ae
 
 
de77829
 
8031785
 
 
 
 
 
 
 
 
 
 
883c794
1acaa19
 
44d964a
c0e541b
95261ed
c0e541b
 
31f7bdb
c0e541b
05a2178
f288ceb
31f7bdb
 
22c7ed2
05a2178
a1da02d
44d964a
 
01fddc0
 
 
20f75ae
 
 
 
 
 
 
 
fdb8dbd
9cae71a
 
a1da02d
 
9cae71a
 
a1da02d
 
b59dd62
 
9cae71a
 
 
a1da02d
9cae71a
8031785
9cae71a
8031785
a1da02d
 
 
 
 
9cae71a
 
fdb8dbd
 
8031785
9cae71a
 
 
 
 
a1da02d
 
 
b59dd62
 
 
9cae71a
b59dd62
 
a1da02d
 
 
b59dd62
 
 
9cae71a
 
 
 
 
 
a1da02d
 
9cae71a
fdb8dbd
 
 
 
 
 
 
8031785
 
a1da02d
 
 
 
 
 
 
8031785
fdb8dbd
 
33a2c1e
f55c594
33a2c1e
fdb8dbd
8031785
f55c594
 
533d92e
cb9ee50
f55c594
fdd892b
 
 
 
2698c96
33ee1bb
 
533d92e
cb9ee50
 
 
 
 
74b1efd
 
fdd892b
cb9ee50
 
22c7ed2
 
c90f138
 
 
 
 
 
 
cb9ee50
 
 
c90f138
cb9ee50
 
 
 
 
 
 
c90f138
 
 
 
 
cb9ee50
8031785
cb9ee50
 
c90f138
 
 
a1da02d
 
 
 
 
 
 
 
 
 
 
 
 
f55c594
cb9ee50
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
71950a8
fdd892b
71950a8
fdd892b
 
883c794
cb9ee50
 
 
 
 
 
 
 
fdd892b
 
 
3fadc6e
8031785
 
c90f138
84fa1f8
 
 
c90f138
 
 
 
84fa1f8
 
 
af09fe7
 
 
 
 
 
 
 
74b7d77
af09fe7
74b7d77
 
 
 
af09fe7
74b7d77
74b1efd
74b7d77
74b1efd
 
8031785
d906b98
8031785
33a2c1e
8031785
d906b98
8031785
33a2c1e
8031785
d906b98
8031785
33a2c1e
8031785
74b1efd
 
5bbbb16
8031785
33a2c1e
95261ed
74b1efd
31f7bdb
 
 
 
 
33a2c1e
31f7bdb
 
33a2c1e
74b1efd
 
 
33a2c1e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
31f7bdb
95261ed
33a2c1e
95261ed
c0e541b
 
 
 
 
 
31f7bdb
c0e541b
31f7bdb
c0e541b
 
 
 
31f7bdb
 
 
c0e541b
4a9d465
 
31f7bdb
 
c0e541b
95261ed
84fa1f8
 
 
 
 
 
 
 
8031785
d906b98
 
 
 
5bbbb16
8031785
 
 
d906b98
5bbbb16
d906b98
f55c594
74b1efd
 
 
 
 
 
 
 
f55c594
 
b84f4f2
74b1efd
 
 
 
 
b84f4f2
74b1efd
 
 
883c794
31f7bdb
74b1efd
883c794
cb9ee50
 
8d120bf
883c794
 
 
 
 
 
 
 
 
f55c594
883c794
6a308c6
883c794
f55c594
883c794
f55c594
883c794
 
3fadc6e
883c794
 
3fadc6e
883c794
 
 
 
7ce6041
883c794
05a2178
31f7bdb
cbc9717
31f7bdb
 
c0e541b
 
 
 
31f7bdb
05a2178
1acaa19
 
 
05a2178
95261ed
1acaa19
 
95261ed
751fc8f
 
 
 
 
 
 
 
 
 
 
 
71950a8
 
05a2178
f5884f3
38cc8a7
751fc8f
 
 
 
1acaa19
 
93c4867
751fc8f
084aa80
44d964a
 
9cae71a
1acaa19
adca588
8d120bf
cb9ee50
71950a8
1acaa19
9cae71a
 
 
1acaa19
 
 
9cae71a
 
 
 
 
fdb8dbd
 
a1da02d
 
 
 
 
 
 
 
 
 
 
b59dd62
 
 
9cae71a
 
 
 
a1da02d
9cae71a
fdb8dbd
 
 
 
 
 
 
b59dd62
 
9cae71a
 
 
 
 
74b7d77
9cae71a
 
 
 
 
fdb8dbd
1acaa19
 
 
 
 
 
 
 
 
 
 
f55c594
a1da02d
 
 
 
 
3fadc6e
8d120bf
3fadc6e
8d120bf
3fadc6e
7ce6041
fdb8dbd
 
d20404a
b59dd62
d20404a
b59dd62
 
 
d20404a
1acaa19
31f7bdb
 
 
05a2178
71950a8
2698c96
 
 
 
 
44d964a
95261ed
2698c96
44d964a
2698c96
44d964a
2698c96
44d964a
2698c96
44d964a
2698c96
d20404a
2698c96
44d964a
2698c96
44d964a
74b7d77
8031785
2698c96
44d964a
2698c96
44d964a
2698c96
44d964a
2698c96
44d964a
2698c96
 
 
 
e694402
33ee1bb
43189ac
 
95261ed
 
1acaa19
2698c96
 
43189ac
 
 
6250a98
1acaa19
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
from datetime import datetime
import json
import math
from typing import Iterator, Union
import argparse

from io import StringIO
import os
import pathlib
import tempfile
import zipfile
import numpy as np

import torch

from src.config import VAD_INITIAL_PROMPT_MODE_VALUES, ApplicationConfig, VadInitialPromptMode
from src.diarization.diarization import Diarization
from src.hooks.progressListener import ProgressListener
from src.hooks.subTaskProgressListener import SubTaskProgressListener
from src.hooks.whisperProgressHook import create_progress_listener_handle
from src.languages import get_language_names
from src.modelCache import ModelCache
from src.prompts.jsonPromptStrategy import JsonPromptStrategy
from src.prompts.prependPromptStrategy import PrependPromptStrategy
from src.source import get_audio_source_collection
from src.vadParallel import ParallelContext, ParallelTranscription

# External programs
import ffmpeg

# UI
import gradio as gr

from src.download import ExceededMaximumDuration, download_url
from src.utils import optional_int, slugify, write_srt, write_vtt
from src.vad import AbstractTranscription, NonSpeechStrategy, PeriodicTranscriptionConfig, TranscriptionConfig, VadPeriodicTranscription, VadSileroTranscription
from src.whisper.abstractWhisperContainer import AbstractWhisperContainer
from src.whisper.whisperFactory import create_whisper_container

# Configure more application defaults in config.json5

# Gradio seems to truncate files without keeping the extension, so we need to truncate the file prefix ourself 
MAX_FILE_PREFIX_LENGTH = 17

# Limit auto_parallel to a certain number of CPUs (specify vad_cpu_cores to get a higher number)
MAX_AUTO_CPU_CORES = 8

WHISPER_MODELS = ["tiny", "base", "small", "medium", "large", "large-v1", "large-v2"]

class VadOptions:
    def __init__(self, vad: str = None, vadMergeWindow: float = 5, vadMaxMergeSize: float = 150, vadPadding: float = 1, vadPromptWindow: float = 1, 
                                        vadInitialPromptMode: Union[VadInitialPromptMode, str] = VadInitialPromptMode.PREPREND_FIRST_SEGMENT):
        self.vad = vad
        self.vadMergeWindow = vadMergeWindow
        self.vadMaxMergeSize = vadMaxMergeSize
        self.vadPadding = vadPadding
        self.vadPromptWindow = vadPromptWindow
        self.vadInitialPromptMode = vadInitialPromptMode if isinstance(vadInitialPromptMode, VadInitialPromptMode) \
                                        else VadInitialPromptMode.from_string(vadInitialPromptMode)

class WhisperTranscriber:
    def __init__(self, input_audio_max_duration: float = None, vad_process_timeout: float = None, 
                 vad_cpu_cores: int = 1, delete_uploaded_files: bool = False, output_dir: str = None, 
                 app_config: ApplicationConfig = None):
        self.model_cache = ModelCache()
        self.parallel_device_list = None
        self.gpu_parallel_context = None
        self.cpu_parallel_context = None
        self.vad_process_timeout = vad_process_timeout
        self.vad_cpu_cores = vad_cpu_cores

        self.vad_model = None
        self.inputAudioMaxDuration = input_audio_max_duration
        self.deleteUploadedFiles = delete_uploaded_files
        self.output_dir = output_dir

        self.diarization: Diarization = None
        self.app_config = app_config

    def set_parallel_devices(self, vad_parallel_devices: str):
        self.parallel_device_list = [ device.strip() for device in vad_parallel_devices.split(",") ] if vad_parallel_devices else None

    def set_auto_parallel(self, auto_parallel: bool):
        if auto_parallel:
            if torch.cuda.is_available():
                self.parallel_device_list = [ str(gpu_id) for gpu_id in range(torch.cuda.device_count())]

            self.vad_cpu_cores = min(os.cpu_count(), MAX_AUTO_CPU_CORES)
            print("[Auto parallel] Using GPU devices " + str(self.parallel_device_list) + " and " + str(self.vad_cpu_cores) + " CPU cores for VAD/transcription.")

    # Entry function for the simple tab
    def transcribe_webui_simple(self, modelName, languageName, urlData, multipleFiles, microphoneData, task, 
                                vad, vadMergeWindow, vadMaxMergeSize, 
                                word_timestamps: bool = False, highlight_words: bool = False,
                                diarization: bool = False, diarization_speakers: int = 2):
        return self.transcribe_webui_simple_progress(modelName, languageName, urlData, multipleFiles, microphoneData, task, 
                                                     vad, vadMergeWindow, vadMaxMergeSize, 
                                                     word_timestamps, highlight_words, 
                                                     diarization, diarization_speakers)
    
    # Entry function for the simple tab progress
    def transcribe_webui_simple_progress(self, modelName, languageName, urlData, multipleFiles, microphoneData, task, 
                                         vad, vadMergeWindow, vadMaxMergeSize, 
                                         word_timestamps: bool = False, highlight_words: bool = False, 
                                         diarization: bool = False, diarization_speakers: int = 2,
                                         progress=gr.Progress()):
        
        vadOptions = VadOptions(vad, vadMergeWindow, vadMaxMergeSize, self.app_config.vad_padding, self.app_config.vad_prompt_window, self.app_config.vad_initial_prompt_mode)

        if diarization:
            self.diarization = Diarization(auth_token=self.app_config.auth_token, num_speakers=diarization_speakers)
        else:
            self.diarization = None

        return self.transcribe_webui(modelName, languageName, urlData, multipleFiles, microphoneData, task, vadOptions, 
                                     word_timestamps=word_timestamps, highlight_words=highlight_words, progress=progress)

    # Entry function for the full tab
    def transcribe_webui_full(self, modelName, languageName, urlData, multipleFiles, microphoneData, task, 
                              vad, vadMergeWindow, vadMaxMergeSize, vadPadding, vadPromptWindow, vadInitialPromptMode, 
                              # Word timestamps
                              word_timestamps: bool, highlight_words: bool, prepend_punctuations: str, append_punctuations: str,
                              initial_prompt: str, temperature: float, best_of: int, beam_size: int, patience: float, length_penalty: float, suppress_tokens: str, 
                              condition_on_previous_text: bool, fp16: bool, temperature_increment_on_fallback: float, 
                              compression_ratio_threshold: float, logprob_threshold: float, no_speech_threshold: float,
                              diarization: bool = False, diarization_speakers: int = 2, 
                              diarization_min_speakers = 1, diarization_max_speakers = 5):
        
        return self.transcribe_webui_full_progress(modelName, languageName, urlData, multipleFiles, microphoneData, task, 
                                vad, vadMergeWindow, vadMaxMergeSize, vadPadding, vadPromptWindow, vadInitialPromptMode,
                                word_timestamps, highlight_words, prepend_punctuations, append_punctuations,
                                initial_prompt, temperature, best_of, beam_size, patience, length_penalty, suppress_tokens,
                                condition_on_previous_text, fp16, temperature_increment_on_fallback,
                                compression_ratio_threshold, logprob_threshold, no_speech_threshold,
                                diarization, diarization_speakers, 
                                diarization_min_speakers, diarization_max_speakers)

    # Entry function for the full tab with progress
    def transcribe_webui_full_progress(self, modelName, languageName, urlData, multipleFiles, microphoneData, task, 
                                        vad, vadMergeWindow, vadMaxMergeSize, vadPadding, vadPromptWindow, vadInitialPromptMode,
                                        # Word timestamps
                                        word_timestamps: bool, highlight_words: bool, prepend_punctuations: str, append_punctuations: str,   
                                        initial_prompt: str, temperature: float, best_of: int, beam_size: int, patience: float, length_penalty: float, suppress_tokens: str, 
                                        condition_on_previous_text: bool, fp16: bool, temperature_increment_on_fallback: float, 
                                        compression_ratio_threshold: float, logprob_threshold: float, no_speech_threshold: float, 
                                        diarization: bool = False, diarization_speakers: int = 2, 
                                        diarization_min_speakers = 1, diarization_max_speakers = 5,
                                        progress=gr.Progress()):

        # Handle temperature_increment_on_fallback
        if temperature_increment_on_fallback is not None:
            temperature = tuple(np.arange(temperature, 1.0 + 1e-6, temperature_increment_on_fallback))
        else:
            temperature = [temperature]

        vadOptions = VadOptions(vad, vadMergeWindow, vadMaxMergeSize, vadPadding, vadPromptWindow, vadInitialPromptMode)

        # Set diarization
        if diarization:
            self.diarization = Diarization(auth_token=self.app_config.auth_token, num_speakers=diarization_speakers, 
                                           min_speakers=diarization_min_speakers, max_speakers=diarization_max_speakers)
        else:
            self.diarization = None

        return self.transcribe_webui(modelName, languageName, urlData, multipleFiles, microphoneData, task, vadOptions,
                                     initial_prompt=initial_prompt, temperature=temperature, best_of=best_of, beam_size=beam_size, patience=patience, length_penalty=length_penalty, suppress_tokens=suppress_tokens,
                                     condition_on_previous_text=condition_on_previous_text, fp16=fp16,
                                     compression_ratio_threshold=compression_ratio_threshold, logprob_threshold=logprob_threshold, no_speech_threshold=no_speech_threshold, 
                                     word_timestamps=word_timestamps, prepend_punctuations=prepend_punctuations, append_punctuations=append_punctuations, highlight_words=highlight_words,
                                     progress=progress)

    def transcribe_webui(self, modelName, languageName, urlData, multipleFiles, microphoneData, task, 
                         vadOptions: VadOptions, progress: gr.Progress = None, highlight_words: bool = False, 
                         **decodeOptions: dict):
        try:
            sources = self.__get_source(urlData, multipleFiles, microphoneData)

            try:
                selectedLanguage = languageName.lower() if len(languageName) > 0 else None
                selectedModel = modelName if modelName is not None else "base"

                model = create_whisper_container(whisper_implementation=self.app_config.whisper_implementation, 
                                                 model_name=selectedModel, compute_type=self.app_config.compute_type, 
                                                 cache=self.model_cache, models=self.app_config.models)

                # Result
                download = []
                zip_file_lookup = {}
                text = ""
                vtt = ""

                # Write result
                downloadDirectory = tempfile.mkdtemp()
                source_index = 0

                outputDirectory = self.output_dir if self.output_dir is not None else downloadDirectory

                # Progress
                total_duration = sum([source.get_audio_duration() for source in sources])
                current_progress = 0

                # A listener that will report progress to Gradio
                root_progress_listener = self._create_progress_listener(progress)

                # Execute whisper
                for source in sources:
                    source_prefix = ""
                    source_audio_duration = source.get_audio_duration()

                    if (len(sources) > 1):
                        # Prefix (minimum 2 digits)
                        source_index += 1
                        source_prefix = str(source_index).zfill(2) + "_"
                        print("Transcribing ", source.source_path)

                    scaled_progress_listener = SubTaskProgressListener(root_progress_listener, 
                                                   base_task_total=total_duration,
                                                   sub_task_start=current_progress,
                                                   sub_task_total=source_audio_duration)

                    # Transcribe
                    result = self.transcribe_file(model, source.source_path, selectedLanguage, task, vadOptions, scaled_progress_listener, **decodeOptions)
                    filePrefix = slugify(source_prefix + source.get_short_name(), allow_unicode=True)

                    # Update progress
                    current_progress += source_audio_duration

                    # Diarization
                    if self.diarization:
                        print("Diarizing ", source.source_path)
                        diarization_result = list(self.diarization.run(source.source_path))

                        # Print result
                        print("Diarization result: ")
                        for entry in diarization_result:
                            print(f"  start={entry.start:.1f}s stop={entry.end:.1f}s speaker_{entry.speaker}")

                        # Add speakers to result
                        result = self.diarization.mark_speakers(diarization_result, result)

                    source_download, source_text, source_vtt = self.write_result(result, filePrefix, outputDirectory, highlight_words)

                    if len(sources) > 1:
                        # Add new line separators
                        if (len(source_text) > 0):
                            source_text += os.linesep + os.linesep
                        if (len(source_vtt) > 0):
                            source_vtt += os.linesep + os.linesep

                        # Append file name to source text too
                        source_text = source.get_full_name() + ":" + os.linesep + source_text
                        source_vtt = source.get_full_name() + ":" + os.linesep + source_vtt

                    # Add to result
                    download.extend(source_download)
                    text += source_text
                    vtt += source_vtt

                    if (len(sources) > 1):
                        # Zip files support at least 260 characters, but we'll play it safe and use 200
                        zipFilePrefix = slugify(source_prefix + source.get_short_name(max_length=200), allow_unicode=True)

                        # File names in ZIP file can be longer
                        for source_download_file in source_download:
                            # Get file postfix (after last -)
                            filePostfix = os.path.basename(source_download_file).split("-")[-1]
                            zip_file_name = zipFilePrefix + "-" + filePostfix
                            zip_file_lookup[source_download_file] = zip_file_name

                # Create zip file from all sources
                if len(sources) > 1:
                    downloadAllPath = os.path.join(downloadDirectory, "All_Output-" + datetime.now().strftime("%Y%m%d-%H%M%S") + ".zip")

                    with zipfile.ZipFile(downloadAllPath, 'w', zipfile.ZIP_DEFLATED) as zip:
                        for download_file in download:
                            # Get file name from lookup
                            zip_file_name = zip_file_lookup.get(download_file, os.path.basename(download_file))
                            zip.write(download_file, arcname=zip_file_name)

                    download.insert(0, downloadAllPath)

                return download, text, vtt

            finally:
                # Cleanup source
                if self.deleteUploadedFiles:
                    for source in sources:
                        print("Deleting source file " + source.source_path)

                        try:
                            os.remove(source.source_path)
                        except Exception as e:
                            # Ignore error - it's just a cleanup
                            print("Error deleting source file " + source.source_path + ": " + str(e))
        
        except ExceededMaximumDuration as e:
            return [], ("[ERROR]: Maximum remote video length is " + str(e.maxDuration) + "s, file was " + str(e.videoDuration) + "s"), "[ERROR]"

    def transcribe_file(self, model: AbstractWhisperContainer, audio_path: str, language: str, task: str = None, 
                        vadOptions: VadOptions = VadOptions(), 
                        progressListener: ProgressListener = None, **decodeOptions: dict):
        
        initial_prompt = decodeOptions.pop('initial_prompt', None)

        if progressListener is None:
            # Default progress listener
            progressListener = ProgressListener()

        if ('task' in decodeOptions):
            task = decodeOptions.pop('task')

        initial_prompt_mode = vadOptions.vadInitialPromptMode

        # Set default initial prompt mode
        if (initial_prompt_mode is None):
            initial_prompt_mode = VadInitialPromptMode.PREPREND_FIRST_SEGMENT

        if (initial_prompt_mode == VadInitialPromptMode.PREPEND_ALL_SEGMENTS or 
            initial_prompt_mode == VadInitialPromptMode.PREPREND_FIRST_SEGMENT):
            # Prepend initial prompt
            prompt_strategy = PrependPromptStrategy(initial_prompt, initial_prompt_mode)
        elif (vadOptions.vadInitialPromptMode == VadInitialPromptMode.JSON_PROMPT_MODE):
            # Use a JSON format to specify the prompt for each segment
            prompt_strategy = JsonPromptStrategy(initial_prompt)
        else:
            raise ValueError("Invalid vadInitialPromptMode: " + initial_prompt_mode)

        # Callable for processing an audio file
        whisperCallable = model.create_callback(language, task, prompt_strategy=prompt_strategy, **decodeOptions)

        # The results
        if (vadOptions.vad == 'silero-vad'):
            # Silero VAD where non-speech gaps are transcribed
            process_gaps = self._create_silero_config(NonSpeechStrategy.CREATE_SEGMENT, vadOptions)
            result = self.process_vad(audio_path, whisperCallable, self.vad_model, process_gaps, progressListener=progressListener)
        elif (vadOptions.vad == 'silero-vad-skip-gaps'):
            # Silero VAD where non-speech gaps are simply ignored
            skip_gaps = self._create_silero_config(NonSpeechStrategy.SKIP, vadOptions)
            result = self.process_vad(audio_path, whisperCallable, self.vad_model, skip_gaps, progressListener=progressListener)
        elif (vadOptions.vad == 'silero-vad-expand-into-gaps'):
            # Use Silero VAD where speech-segments are expanded into non-speech gaps
            expand_gaps = self._create_silero_config(NonSpeechStrategy.EXPAND_SEGMENT, vadOptions)
            result = self.process_vad(audio_path, whisperCallable, self.vad_model, expand_gaps, progressListener=progressListener)
        elif (vadOptions.vad == 'periodic-vad'):
            # Very simple VAD - mark every 5 minutes as speech. This makes it less likely that Whisper enters an infinite loop, but
            # it may create a break in the middle of a sentence, causing some artifacts.
            periodic_vad = VadPeriodicTranscription()
            period_config = PeriodicTranscriptionConfig(periodic_duration=vadOptions.vadMaxMergeSize, max_prompt_window=vadOptions.vadPromptWindow)
            result = self.process_vad(audio_path, whisperCallable, periodic_vad, period_config, progressListener=progressListener)

        else:
            if (self._has_parallel_devices()):
                # Use a simple period transcription instead, as we need to use the parallel context
                periodic_vad = VadPeriodicTranscription()
                period_config = PeriodicTranscriptionConfig(periodic_duration=math.inf, max_prompt_window=1)

                result = self.process_vad(audio_path, whisperCallable, periodic_vad, period_config, progressListener=progressListener)
            else:
                # Default VAD
                result = whisperCallable.invoke(audio_path, 0, None, None, progress_listener=progressListener)

        return result

    def _create_progress_listener(self, progress: gr.Progress):
        if (progress is None):
            # Dummy progress listener
            return ProgressListener()
        
        class ForwardingProgressListener(ProgressListener):
            def __init__(self, progress: gr.Progress):
                self.progress = progress

            def on_progress(self, current: Union[int, float], total: Union[int, float]):
                # From 0 to 1
                self.progress(current / total)

            def on_finished(self):
                self.progress(1)

        return ForwardingProgressListener(progress)

    def process_vad(self, audio_path, whisperCallable, vadModel: AbstractTranscription, vadConfig: TranscriptionConfig, 
                    progressListener: ProgressListener = None):
        if (not self._has_parallel_devices()):
            # No parallel devices, so just run the VAD and Whisper in sequence
            return vadModel.transcribe(audio_path, whisperCallable, vadConfig, progressListener=progressListener)

        gpu_devices = self.parallel_device_list

        if (gpu_devices is None or len(gpu_devices) == 0):
            # No GPU devices specified, pass the current environment variable to the first GPU process. This may be NULL.
            gpu_devices = [os.environ.get("CUDA_VISIBLE_DEVICES", None)]

        # Create parallel context if needed
        if (self.gpu_parallel_context is None):
            # Create a context wih processes and automatically clear the pool after 1 hour of inactivity
            self.gpu_parallel_context = ParallelContext(num_processes=len(gpu_devices), auto_cleanup_timeout_seconds=self.vad_process_timeout)
        # We also need a CPU context for the VAD
        if (self.cpu_parallel_context is None):
            self.cpu_parallel_context = ParallelContext(num_processes=self.vad_cpu_cores, auto_cleanup_timeout_seconds=self.vad_process_timeout)

        parallel_vad = ParallelTranscription()
        return parallel_vad.transcribe_parallel(transcription=vadModel, audio=audio_path, whisperCallable=whisperCallable,  
                                                config=vadConfig, cpu_device_count=self.vad_cpu_cores, gpu_devices=gpu_devices, 
                                                cpu_parallel_context=self.cpu_parallel_context, gpu_parallel_context=self.gpu_parallel_context, 
                                                progress_listener=progressListener) 

    def _has_parallel_devices(self):
        return (self.parallel_device_list is not None and len(self.parallel_device_list) > 0) or self.vad_cpu_cores > 1

    def _concat_prompt(self, prompt1, prompt2):
        if (prompt1 is None):
            return prompt2
        elif (prompt2 is None):
            return prompt1
        else:
            return prompt1 + " " + prompt2

    def _create_silero_config(self, non_speech_strategy: NonSpeechStrategy, vadOptions: VadOptions):
        # Use Silero VAD 
        if (self.vad_model is None):
            self.vad_model = VadSileroTranscription()

        config = TranscriptionConfig(non_speech_strategy = non_speech_strategy, 
                max_silent_period=vadOptions.vadMergeWindow, max_merge_size=vadOptions.vadMaxMergeSize, 
                segment_padding_left=vadOptions.vadPadding, segment_padding_right=vadOptions.vadPadding, 
                max_prompt_window=vadOptions.vadPromptWindow)

        return config

    def write_result(self, result: dict, source_name: str, output_dir: str, highlight_words: bool = False):
        if not os.path.exists(output_dir):
            os.makedirs(output_dir)

        text = result["text"]
        language = result["language"]
        languageMaxLineWidth = self.__get_max_line_width(language)

        print("Max line width " + str(languageMaxLineWidth))
        vtt = self.__get_subs(result["segments"], "vtt", languageMaxLineWidth, highlight_words=highlight_words)
        srt = self.__get_subs(result["segments"], "srt", languageMaxLineWidth, highlight_words=highlight_words)
        json_result = json.dumps(result, indent=4, ensure_ascii=False)

        output_files = []
        output_files.append(self.__create_file(srt, output_dir, source_name + "-subs.srt"));
        output_files.append(self.__create_file(vtt, output_dir, source_name + "-subs.vtt"));
        output_files.append(self.__create_file(text, output_dir, source_name + "-transcript.txt"));
        output_files.append(self.__create_file(json_result, output_dir, source_name + "-result.json"));

        return output_files, text, vtt

    def clear_cache(self):
        self.model_cache.clear()
        self.vad_model = None

    def __get_source(self, urlData, multipleFiles, microphoneData):
        return get_audio_source_collection(urlData, multipleFiles, microphoneData, self.inputAudioMaxDuration)

    def __get_max_line_width(self, language: str) -> int:
        if (language and language.lower() in ["japanese", "ja", "chinese", "zh"]):
            # Chinese characters and kana are wider, so limit line length to 40 characters
            return 40
        else:
            # TODO: Add more languages
            # 80 latin characters should fit on a 1080p/720p screen
            return 80

    def __get_subs(self, segments: Iterator[dict], format: str, maxLineWidth: int, highlight_words: bool = False) -> str:
        segmentStream = StringIO()

        if format == 'vtt':
            write_vtt(segments, file=segmentStream, maxLineWidth=maxLineWidth, highlight_words=highlight_words)
        elif format == 'srt':
            write_srt(segments, file=segmentStream, maxLineWidth=maxLineWidth, highlight_words=highlight_words)
        else:
            raise Exception("Unknown format " + format)

        segmentStream.seek(0)
        return segmentStream.read()

    def __create_file(self, text: str, directory: str, fileName: str) -> str:
        # Write the text to a file
        with open(os.path.join(directory, fileName), 'w+', encoding="utf-8") as file:
            file.write(text)

        return file.name

    def close(self):
        print("Closing parallel contexts")
        self.clear_cache()

        if (self.gpu_parallel_context is not None):
            self.gpu_parallel_context.close()
        if (self.cpu_parallel_context is not None):
            self.cpu_parallel_context.close()


def create_ui(app_config: ApplicationConfig):
    ui = WhisperTranscriber(app_config.input_audio_max_duration, app_config.vad_process_timeout, app_config.vad_cpu_cores, 
                            app_config.delete_uploaded_files, app_config.output_dir, app_config)

    # Specify a list of devices to use for parallel processing
    ui.set_parallel_devices(app_config.vad_parallel_devices)
    ui.set_auto_parallel(app_config.auto_parallel)

    is_whisper = False

    if app_config.whisper_implementation == "whisper":
        implementation_name = "Whisper"
        is_whisper = True
    elif app_config.whisper_implementation in ["faster-whisper", "faster_whisper"]:
        implementation_name = "Faster Whisper"
    else:
        # Try to convert from camel-case to title-case
        implementation_name = app_config.whisper_implementation.title().replace("_", " ").replace("-", " ")

    ui_description = implementation_name + " is a general-purpose speech recognition model. It is trained on a large dataset of diverse " 
    ui_description += " audio and is also a multi-task model that can perform multilingual speech recognition "
    ui_description += " as well as speech translation and language identification. "

    ui_description += "\n\n\n\nFor longer audio files (>10 minutes) not in English, it is recommended that you select Silero VAD (Voice Activity Detector) in the VAD option."

    # Recommend faster-whisper
    if is_whisper:
        ui_description += "\n\n\n\nFor faster inference on GPU, try [faster-whisper](https://huggingface.co/spaces/aadnk/faster-whisper-webui)."

    if app_config.input_audio_max_duration > 0:
        ui_description += "\n\n" + "Max audio file length: " + str(app_config.input_audio_max_duration) + " s"

    ui_article = "Read the [documentation here](https://gitlab.com/aadnk/whisper-webui/-/blob/main/docs/options.md)."

    whisper_models = app_config.get_model_names()

    common_inputs = lambda : [
        gr.Dropdown(choices=whisper_models, value=app_config.default_model_name, label="Model"),
        gr.Dropdown(choices=sorted(get_language_names()), label="Language", value=app_config.language),
        gr.Text(label="URL (YouTube, etc.)"),
        gr.File(label="Upload Files", file_count="multiple"),
        gr.Audio(source="microphone", type="filepath", label="Microphone Input"),
        gr.Dropdown(choices=["transcribe", "translate"], label="Task", value=app_config.task),
    ]

    common_vad_inputs = lambda : [
        gr.Dropdown(choices=["none", "silero-vad", "silero-vad-skip-gaps", "silero-vad-expand-into-gaps", "periodic-vad"], value=app_config.default_vad, label="VAD"),
        gr.Number(label="VAD - Merge Window (s)", precision=0, value=app_config.vad_merge_window),
        gr.Number(label="VAD - Max Merge Size (s)", precision=0, value=app_config.vad_max_merge_size),
    ]
    
    common_word_timestamps_inputs = lambda : [
        gr.Checkbox(label="Word Timestamps", value=app_config.word_timestamps),
        gr.Checkbox(label="Word Timestamps - Highlight Words", value=app_config.highlight_words),
    ]

    has_diarization_libs = Diarization.has_libraries()

    if not has_diarization_libs:
        print("Diarization libraries not found - disabling diarization")
        app_config.diarization = False

    common_diarization_inputs = lambda : [
        gr.Checkbox(label="Diarization", value=app_config.diarization, interactive=has_diarization_libs),
        gr.Number(label="Diarization - Speakers", precision=0, value=app_config.diarization_speakers, interactive=has_diarization_libs)
    ]

    is_queue_mode = app_config.queue_concurrency_count is not None and app_config.queue_concurrency_count > 0    

    simple_transcribe = gr.Interface(fn=ui.transcribe_webui_simple_progress if is_queue_mode else ui.transcribe_webui_simple, 
                                     description=ui_description, article=ui_article, inputs=[
        *common_inputs(),
        *common_vad_inputs(),
        *common_word_timestamps_inputs(),
        *common_diarization_inputs(),
    ], outputs=[
        gr.File(label="Download"),
        gr.Text(label="Transcription"), 
        gr.Text(label="Segments")
    ])

    full_description = ui_description + "\n\n\n\n" + "Be careful when changing some of the options in the full interface - this can cause the model to crash."

    full_transcribe = gr.Interface(fn=ui.transcribe_webui_full_progress if is_queue_mode else ui.transcribe_webui_full,
                                   description=full_description, article=ui_article, inputs=[
        *common_inputs(),

        *common_vad_inputs(),
        gr.Number(label="VAD - Padding (s)", precision=None, value=app_config.vad_padding),
        gr.Number(label="VAD - Prompt Window (s)", precision=None, value=app_config.vad_prompt_window),
        gr.Dropdown(choices=VAD_INITIAL_PROMPT_MODE_VALUES, label="VAD - Initial Prompt Mode"),
        
        *common_word_timestamps_inputs(),
        gr.Text(label="Word Timestamps - Prepend Punctuations", value=app_config.prepend_punctuations),
        gr.Text(label="Word Timestamps - Append Punctuations", value=app_config.append_punctuations),

        gr.TextArea(label="Initial Prompt"),
        gr.Number(label="Temperature", value=app_config.temperature),
        gr.Number(label="Best Of - Non-zero temperature", value=app_config.best_of, precision=0),
        gr.Number(label="Beam Size - Zero temperature", value=app_config.beam_size, precision=0),
        gr.Number(label="Patience - Zero temperature", value=app_config.patience),
        gr.Number(label="Length Penalty - Any temperature", value=app_config.length_penalty), 
        gr.Text(label="Suppress Tokens - Comma-separated list of token IDs", value=app_config.suppress_tokens),
        gr.Checkbox(label="Condition on previous text", value=app_config.condition_on_previous_text),
        gr.Checkbox(label="FP16", value=app_config.fp16),
        gr.Number(label="Temperature increment on fallback", value=app_config.temperature_increment_on_fallback),
        gr.Number(label="Compression ratio threshold", value=app_config.compression_ratio_threshold),
        gr.Number(label="Logprob threshold", value=app_config.logprob_threshold),
        gr.Number(label="No speech threshold", value=app_config.no_speech_threshold),

        *common_diarization_inputs(),
        gr.Number(label="Diarization - Min Speakers", precision=0, value=app_config.diarization_min_speakers, interactive=has_diarization_libs),
        gr.Number(label="Diarization - Max Speakers", precision=0, value=app_config.diarization_max_speakers, interactive=has_diarization_libs),

    ], outputs=[
        gr.File(label="Download"),
        gr.Text(label="Transcription"), 
        gr.Text(label="Segments")
    ])

    demo = gr.TabbedInterface([simple_transcribe, full_transcribe], tab_names=["Simple", "Full"])

    # Queue up the demo
    if is_queue_mode:
        demo.queue(concurrency_count=app_config.queue_concurrency_count)
        print("Queue mode enabled (concurrency count: " + str(app_config.queue_concurrency_count) + ")")
    else:
        print("Queue mode disabled - progress bars will not be shown.")
   
    demo.launch(share=app_config.share, server_name=app_config.server_name, server_port=app_config.server_port)
    
    # Clean up
    ui.close()

if __name__ == '__main__':
    default_app_config = ApplicationConfig.create_default()
    whisper_models = default_app_config.get_model_names()

    # Environment variable overrides
    default_whisper_implementation = os.environ.get("WHISPER_IMPLEMENTATION", default_app_config.whisper_implementation)

    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
    parser.add_argument("--input_audio_max_duration", type=int, default=default_app_config.input_audio_max_duration, \
                        help="Maximum audio file length in seconds, or -1 for no limit.") # 600
    parser.add_argument("--share", type=bool, default=default_app_config.share, \
                        help="True to share the app on HuggingFace.") # False
    parser.add_argument("--server_name", type=str, default=default_app_config.server_name, \
                        help="The host or IP to bind to. If None, bind to localhost.") # None
    parser.add_argument("--server_port", type=int, default=default_app_config.server_port, \
                        help="The port to bind to.") # 7860
    parser.add_argument("--queue_concurrency_count", type=int, default=default_app_config.queue_concurrency_count, \
                        help="The number of concurrent requests to process.") # 1
    parser.add_argument("--default_model_name", type=str, choices=whisper_models, default=default_app_config.default_model_name, \
                        help="The default model name.") # medium
    parser.add_argument("--default_vad", type=str, default=default_app_config.default_vad, \
                        help="The default VAD.") # silero-vad
    parser.add_argument("--vad_initial_prompt_mode", type=str, default=default_app_config.vad_initial_prompt_mode, choices=VAD_INITIAL_PROMPT_MODE_VALUES, \
                        help="Whether or not to prepend the initial prompt to each VAD segment (prepend_all_segments), or just the first segment (prepend_first_segment)") # prepend_first_segment
    parser.add_argument("--vad_parallel_devices", type=str, default=default_app_config.vad_parallel_devices, \
                        help="A commma delimited list of CUDA devices to use for parallel processing. If None, disable parallel processing.") # ""
    parser.add_argument("--vad_cpu_cores", type=int, default=default_app_config.vad_cpu_cores, \
                        help="The number of CPU cores to use for VAD pre-processing.") # 1
    parser.add_argument("--vad_process_timeout", type=float, default=default_app_config.vad_process_timeout, \
                        help="The number of seconds before inactivate processes are terminated. Use 0 to close processes immediately, or None for no timeout.") # 1800
    parser.add_argument("--auto_parallel", type=bool, default=default_app_config.auto_parallel, \
                        help="True to use all available GPUs and CPU cores for processing. Use vad_cpu_cores/vad_parallel_devices to specify the number of CPU cores/GPUs to use.") # False
    parser.add_argument("--output_dir", "-o", type=str, default=default_app_config.output_dir, \
                        help="directory to save the outputs")
    parser.add_argument("--whisper_implementation", type=str, default=default_whisper_implementation, choices=["whisper", "faster-whisper"],\
                        help="the Whisper implementation to use")
    parser.add_argument("--compute_type", type=str, default=default_app_config.compute_type, choices=["default", "auto", "int8", "int8_float16", "int16", "float16", "float32"], \
                        help="the compute type to use for inference")
    parser.add_argument("--threads", type=optional_int, default=0, 
                        help="number of threads used by torch for CPU inference; supercedes MKL_NUM_THREADS/OMP_NUM_THREADS")

    args = parser.parse_args().__dict__

    updated_config = default_app_config.update(**args)

    if (threads := args.pop("threads")) > 0:
        torch.set_num_threads(threads)

    print("Using whisper implementation: " + updated_config.whisper_implementation)
    create_ui(app_config=updated_config)