added api_name for generation, added whisperx backend, relocated use whispercpp option to whisper backend list

requirements.txt

@@ -1,4 +1,6 @@
 git+https://github.com/openai/whisper.git
+git+https://github.com/m-bain/whisperx.git
+
 more-itertools
 ffmpeg-python
 gradio

src/utils.py

@@ -37,8 +37,11 @@ from tortoise.utils.text import split_and_recombine_text
 from tortoise.utils.device import get_device_name, set_device_name
 
 MODELS['dvae.pth'] = "https://huggingface.co/jbetker/tortoise-tts-v2/resolve/3704aea61678e7e468a06d8eea121dba368a798e/.models/dvae.pth"
-WHISPER_MODELS = ["tiny", "base", "small", "medium", "large"]
+
+WHISPER_MODELS = ["tiny", "base", "small", "medium", "large", "large-v2"]
 WHISPER_SPECIALIZED_MODELS = ["tiny.en", "base.en", "small.en", "medium.en"]
+WHISPER_BACKENDS = ["openai/whisper", "lightmare/whispercpp", "m-bain/whisperx"]
+
 EPOCH_SCHEDULE = [ 9, 18, 25, 33 ]
 
 args = None
@@ -1001,12 +1004,13 @@ def whisper_transcribe( file, language=None ):
 	if not whisper_model:
 		load_whisper_model(language=language)
 
-	if not args.whisper_cpp:
+	if args.whisper_backend == "openai/whisper":
 		if not language:
 			language = None
 
 		return whisper_model.transcribe(file, language=language)
 
+	elif args.whisper_backend == "lightmare/whispercpp":
 		res = whisper_model.transcribe(file)
 		segments = whisper_model.extract_text_and_timestamps( res )
 
@@ -1023,6 +1027,15 @@ def whisper_transcribe( file, language=None ):
 
 		return result
 
+	# credit to https://git.ecker.tech/yqxtqymn for the busywork of getting this added
+	elif args.whisper_backend == "m-bain/whisperx":
+		import whisperx
+		device = "cuda" if get_device_name() == "cuda" else "cpu"
+		result = whisper_model.transcribe(file)
+		model_a, metadata = whisperx.load_align_model(language_code=result["language"], device=device)
+		result_aligned = whisperx.align(result["segments"], model_a, metadata, file, device)
+
+		return result
 
 def prepare_dataset( files, outdir, language=None, progress=None ):
 	unload_tts()
@@ -1425,8 +1438,8 @@ def setup_args():
 		'output-volume': 1,
 		
 		'autoregressive-model': None,
+		'whisper-backend': 'openai/whisper',
 		'whisper-model': "base",
-		'whisper-cpp': False,
 
 		'training-default-halfp': False,
 		'training-default-bnb': True,
@@ -1464,8 +1477,8 @@ def setup_args():
 	parser.add_argument("--output-volume", type=float, default=default_arguments['output-volume'], help="Adjusts volume of output")
 	
 	parser.add_argument("--autoregressive-model", default=default_arguments['autoregressive-model'], help="Specifies which autoregressive model to use for sampling.")
+	parser.add_argument("--whisper-backend", default=default_arguments['whisper-backend'], action='store_true', help="Picks which whisper backend to use (openai/whisper, lightmare/whispercpp, m-bain/whisperx)")
 	parser.add_argument("--whisper-model", default=default_arguments['whisper-model'], help="Specifies which whisper model to use for transcription.")
-	parser.add_argument("--whisper-cpp", default=default_arguments['whisper-cpp'], action='store_true', help="Leverages lightmare/whispercpp for transcription")
 	
 	parser.add_argument("--training-default-halfp", action='store_true', default=default_arguments['training-default-halfp'], help="Training default: halfp")
 	parser.add_argument("--training-default-bnb", action='store_true', default=default_arguments['training-default-bnb'], help="Training default: bnb")
@@ -1499,7 +1512,7 @@ def setup_args():
 	
 	return args
 
-def update_args( listen, share, check_for_updates, models_from_local_only, low_vram, embed_output_metadata, latents_lean_and_mean, voice_fixer, voice_fixer_use_cuda, force_cpu_for_conditioning_latents, defer_tts_load, prune_nonfinal_outputs, use_bigvgan_vocoder, device_override, sample_batch_size, concurrency_count, autocalculate_voice_chunk_duration_size, output_volume, autoregressive_model, whisper_model, whisper_cpp, training_default_halfp, training_default_bnb ):
+def update_args( listen, share, check_for_updates, models_from_local_only, low_vram, embed_output_metadata, latents_lean_and_mean, voice_fixer, voice_fixer_use_cuda, force_cpu_for_conditioning_latents, defer_tts_load, prune_nonfinal_outputs, use_bigvgan_vocoder, device_override, sample_batch_size, concurrency_count, autocalculate_voice_chunk_duration_size, output_volume, autoregressive_model, whisper_backend, whisper_model, training_default_halfp, training_default_bnb ):
 	global args
 
 	args.listen = listen
@@ -1523,8 +1536,8 @@ def update_args( listen, share, check_for_updates, models_from_local_only, low_v
 	args.output_volume = output_volume
 	
 	args.autoregressive_model = autoregressive_model
+	args.whisper_backend = whisper_backend
 	args.whisper_model = whisper_model
-	args.whisper_cpp = whisper_cpp
 
 	args.training_default_halfp = training_default_halfp
 	args.training_default_bnb = training_default_bnb
@@ -1555,8 +1568,8 @@ def save_args_settings():
 		'output-volume': args.output_volume,
 		
 		'autoregressive-model': args.autoregressive_model,
+		'whisper-backend': args.whisper_backend,
 		'whisper-model': args.whisper_model,
-		'whisper-cpp': args.whisper_cpp,
 
 		'training-default-halfp': args.training_default_halfp,
 		'training-default-bnb': args.training_default_bnb,
@@ -1762,6 +1775,12 @@ def unload_voicefixer():
 def load_whisper_model(language=None, model_name=None, progress=None):
 	global whisper_model
 
+	if args.whisper_backend not in WHISPER_BACKENDS:
+		raise Exception(f"unavailable backend: {args.whisper_backend}")
+
+	if args.whisper_backend != "m-bain/whisperx" and model_name == "large-v2":
+		raise Exception("large-v2 is only available for m-bain/whisperx backend")
+	
 	if not model_name:
 		model_name = args.whisper_model
 	else:
@@ -1774,16 +1793,20 @@ def load_whisper_model(language=None, model_name=None, progress=None):
 
 	notify_progress(f"Loading Whisper model: {model_name}", progress)
 
-	if args.whisper_cpp:
+	if args.whisper_backend == "openai/whisper":
+		import whisper
+		whisper_model = whisper.load_model(model_name)
+	elif args.whisper_backend == "lightmare/whispercpp":
 		from whispercpp import Whisper
 		if not language:
 			language = 'auto'
 
 		b_lang = language.encode('ascii')
 		whisper_model = Whisper(model_name, models_dir='./models/', language=b_lang)
-	else:
+	elif args.whisper_backend == "m-bain/whisperx":
-		import whisper
+		import whisperx
-		whisper_model = whisper.load_model(model_name)
+		device = "cuda" if get_device_name() == "cuda" else "cpu"
+		whisper_model = whisperx.load_model(model_name, device)
 
 	print("Loaded Whisper model")
 

src/webui.py

@@ -589,10 +589,10 @@ def setup_gradio():
 					
 					autoregressive_model_dropdown = gr.Dropdown(choices=autoregressive_models, label="Autoregressive Model", value=args.autoregressive_model if args.autoregressive_model else autoregressive_models[0])
 					
+					whisper_backend = gr.Dropdown(WHISPER_BACKENDS, label="Whisper Backends", value=args.whisper_backend)
 					whisper_model_dropdown = gr.Dropdown(WHISPER_MODELS, label="Whisper Model", value=args.whisper_model)
-					use_whisper_cpp = gr.Checkbox(label="Use Whisper.cpp", value=args.whisper_cpp)
 					
-					exec_inputs = exec_inputs + [ autoregressive_model_dropdown, whisper_model_dropdown, use_whisper_cpp, training_halfp, training_bnb ]
+					exec_inputs = exec_inputs + [ autoregressive_model_dropdown, whisper_backend, whisper_model_dropdown, training_halfp, training_bnb ]
 
 					with gr.Row():
 						autoregressive_models_update_button = gr.Button(value="Refresh Model List")
@@ -732,6 +732,7 @@ def setup_gradio():
 		submit_event = submit.click(run_generation,
 			inputs=input_settings,
 			outputs=[output_audio, source_sample, candidates_list, generation_results],
+			api_name="generate",
 		)
 
 

update.sh

@@ -3,5 +3,6 @@ git pull
 git submodule update --remote
 
 if python -m pip show whispercpp &>/dev/null; then python -m pip install -U git+https://git.ecker.tech/lightmare/whispercpp.py; fi
+if python -m pip show whisperx &>/dev/null; then python -m pip install -U git+https://github.com/m-bain/whisperx.git; fi
 
 deactivate