Added experimental guided setup

requirements.txt

@@ -1,4 +1,3 @@
-git+https://github.com/openai/whisper.git
 more-itertools
 ffmpeg-python
 gradio

setup-cuda.bat

@@ -5,10 +5,10 @@ python -m venv venv
 call .\venv\Scripts\activate.bat
 python -m pip install --upgrade pip
 python -m pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117
-python -m pip install -r .\dlas\requirements.txt
+python -m pip install -r .\requirements.txt
 python -m pip install -r .\tortoise-tts\requirements.txt
 python -m pip install -e .\tortoise-tts\
-python -m pip install -r .\requirements.txt
+python -m pip install -r .\dlas\requirements.txt
 
 xcopy .\dlas\bitsandbytes_windows\* .\venv\Lib\site-packages\bitsandbytes\. /Y
 xcopy .\dlas\bitsandbytes_windows\cuda_setup\* .\venv\Lib\site-packages\bitsandbytes\cuda_setup\. /Y

setup-cuda.sh

@@ -9,9 +9,9 @@ python3 -m pip install --upgrade pip # just to be safe
 # CUDA
 pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117
 # install requirements
-python3 -m pip install -r ./dlas/requirements.txt # instal DLAS requirements
+python3 -m pip install -r ./requirements.txt # install local requirements
 python3 -m pip install -r ./tortoise-tts/requirements.txt # install TorToiSe requirements
 python3 -m pip install -e ./tortoise-tts/ # install TorToiSe
-python3 -m pip install -r ./requirements.txt # install local requirements
+python3 -m pip install -r ./dlas/requirements.txt # instal DLAS requirements, last, because whisperx will break a dependency here
 
 deactivate

setup-directml.bat

@@ -5,10 +5,10 @@ python -m venv venv
 call .\venv\Scripts\activate.bat
 python -m pip install --upgrade pip
 python -m pip install torch torchvision torchaudio torch-directml
-python -m pip install -r .\dlas\requirements.txt
+python -m pip install -r .\requirements.txt
 python -m pip install -r .\tortoise-tts\requirements.txt
 python -m pip install -e .\tortoise-tts\
-python -m pip install -r .\requirements.txt
+python -m pip install -r .\dlas\requirements.txt
 
 pause
 deactivate

setup-guided.sh

@@ -0,0 +1,54 @@
+#!/bin/bash
+
+if ! command -v git &> /dev/null; then
+	if [[ "$(read -e -p 'Could not find git. Continue? [y/N]> '; echo $REPLY)" != [Yy]* ]]; then exit 1; fi
+else
+	printf "git - ok\n"
+fi
+
+# TODO: This could be more precise. e.g. checking for python3, then python, etc.
+if ! command -v python &> /dev/null; then
+	if [[ "$(read -e -p 'Could not find python. Continue? [y/N]> '; echo $REPLY)" != [Yy]* ]]; then exit 1; fi
+else
+	python -c 'import sys; sys.stderr.write("Wrong python version.\n") if sys.version_info.major != 3 else sys.stderr.write("Python 3 - ok\n")'
+fi
+
+printf "Which GPU brand do you have?\n"
+gpus=(nvidia amd)
+
+gpu=""
+while [ "$gpu" = "" ]; do
+	select gpu in $(printf '%s\n' ${gpus[@]}); do break; done
+done
+
+if [ $gpu = "nvidia" ]; then
+	./setup-cuda.sh
+elif [ $gpu = "amd" ]; then
+	./setup-rocm.sh
+fi
+
+source ./venv/bin/activate
+
+printf "Which Whisper backend would you like to use?\n"
+whisper_backends=("openai/whisper" "m-bain/whisperx" "lightmare/whispercpp")
+
+whisper_backend=""
+while [ "$whisper_backend" = "" ]; do
+	select whisper_backend in $(printf '%s\n' ${whisper_backends[@]}); do break; done
+done
+
+if [ $whisper_backend = "openai/whisper" ]; then
+	python -m pip install git+https://github.com/openai/whisper.git
+elif [ $whisper_backend = "m-bain/whisperx" ]; then
+	python -m pip install git+https://github.com/m-bain/whisperx.git
+elif [ $whisper_backend = "lightmare/whispercpp" ]; then
+	# This depends on SemVer
+	# Git > v2.18 for `--sort`
+	# Git > v2.4 for `versionsort.suffix`
+	# For older versions:
+	# git ls-remote --refs --tags https://git.ecker.tech/lightmare/whispercpp.py | cut --delimiter='/' --fields=3 | tr '-' '~' | sort --version-sort | tail --lines=1
+	WHISPERCPP_LATEST=$(git -c 'versionsort.suffix=-' ls-remote --exit-code --refs --sort='version:refname' --tags https://git.ecker.tech/lightmare/whispercpp.py '*.*.*' | tail -n 1 | cut --delimiter='/' --fields=3)
+	python -m pip install git+https://git.ecker.tech/lightmare/whispercpp.py@$WHISPERCPP_LATEST
+fi
+
+deactivate

setup-rocm.sh

@@ -9,10 +9,10 @@ python3 -m pip install --upgrade pip # just to be safe
 # ROCM
 pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/rocm5.1.1 # 5.2 does not work for me desu
 # install requirements
-python3 -m pip install -r ./dlas/requirements.txt # instal DLAS requirements
+python3 -m pip install -r ./requirements.txt # install local requirements
 python3 -m pip install -r ./tortoise-tts/requirements.txt # install TorToiSe requirements
 python3 -m pip install -e ./tortoise-tts/ # install TorToiSe
-python3 -m pip install -r ./requirements.txt # install local requirements
+python3 -m pip install -r ./dlas/requirements.txt # instal DLAS requirements
 # swap to ROCm version of BitsAndBytes
 pip3 uninstall -y bitsandbytes
 ./setup-rocm-bnb.sh

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
@@ -233,7 +236,7 @@ def generate(
 		if emotion == "Custom":
 			if prompt and prompt.strip() != "":
 				cut_text = f"[{prompt},] {cut_text}"
-		else:
+		elif emotion != "None":
 			cut_text = f"[I am really {emotion.lower()},] {cut_text}"
 
 		progress.msg_prefix = f'[{str(line+1)}/{str(len(texts))}]'
@@ -464,14 +467,21 @@ def update_baseline_for_latents_chunks( voice ):
 		return 1
 
 	files = os.listdir(path)
+	
+	total = 0
 	total_duration = 0
+
 	for file in files:
 		if file[-4:] != ".wav":
 			continue
+
 		metadata = torchaudio.info(f'{path}/{file}')
 		duration = metadata.num_channels * metadata.num_frames / metadata.sample_rate
 		total_duration += duration
+		total = total + 1
 
+	if args.autocalculate_voice_chunk_duration_size == 0:
+		return int(total_duration / total) if total > 0 else 1
 	return int(total_duration / args.autocalculate_voice_chunk_duration_size) if total_duration > 0 else 1
 
 def compute_latents(voice, voice_latents_chunks, progress=gr.Progress(track_tqdm=True)):
@@ -550,6 +560,8 @@ class TrainingState():
 		self.eta = "?"
 		self.eta_hhmmss = "?"
 
+		self.nan_detected = False
+
 		self.last_info_check_at = 0
 		self.statistics = []
 		self.losses = []
@@ -701,13 +713,10 @@ class TrainingState():
 				info_line = line.split("INFO:")[-1]
 				# to-do, actually validate this works, and probably kill training when it's found, the model's dead by this point
 				if ': nan' in info_line:
-					should_return = True
-
-					print("! NAN DETECTED !")
-					self.buffer.append("! NAN DETECTED !")
+					self.nan_detected = True
 
 				# easily rip out our stats...
-				match = re.findall(r'\b([a-z_0-9]+?)\b: +?([0-9]\.[0-9]+?e[+-]\d+|[\d,]+)\b', info_line)
+				match = re.findall(r'\b([a-z_0-9]+?)\b: *?([0-9]\.[0-9]+?e[+-]\d+|[\d,]+)\b', info_line)
 				if match and len(match) > 0:
 					for k, v in match:
 						self.info[k] = float(v.replace(",", ""))
@@ -862,6 +871,8 @@ class TrainingState():
 			self.metrics['loss'] = ", ".join(self.metrics['loss'])
 
 			message = f"[{self.metrics['step']}] [{self.metrics['rate']}] [ETA: {eta_hhmmss}]\n[{self.metrics['loss']}]"
+			if self.nan_detected:
+				message = f"[!NaN DETECTED!] {message}"
 
 			if message:
 				percent = self.it / float(self.its) # self.epoch / float(self.epochs)
@@ -916,12 +927,6 @@ def run_training(config_path, verbose=False, gpus=1, keep_x_past_datasets=0, pro
 		return_code = training_state.process.wait()
 		training_state = None
 
-def get_training_losses():
-	global training_state
-	if not training_state or not training_state.statistics:
-		return
-	return pd.DataFrame(training_state.statistics)
-
 def update_training_dataplot(config_path=None):
 	global training_state
 	update = None
@@ -930,12 +935,12 @@ def update_training_dataplot(config_path=None):
 		if config_path:
 			training_state = TrainingState(config_path=config_path, start=False)
 			if training_state.statistics:
-				update = gr.LinePlot.update(value=pd.DataFrame(training_state.statistics))
+				update = gr.LinePlot.update(value=pd.DataFrame(training_state.statistics), x_lim=[0,training_state.its], x="step", y="value", title="Training Metrics", color="type", tooltip=['step', 'value', 'type'], width=600, height=350,)
 			del training_state
 			training_state = None
 	elif training_state.statistics:
 		training_state.load_losses()
-		update = gr.LinePlot.update(value=pd.DataFrame(training_state.statistics))
+		update = gr.LinePlot.update(value=pd.DataFrame(training_state.statistics), x_lim=[0,training_state.its], x="step", y="value", title="Training Metrics", color="type", tooltip=['step', 'value', 'type'], width=600, height=350,)
 
 	return update
 
@@ -965,7 +970,6 @@ def stop_training():
 	try:
 		children = [p.info for p in psutil.process_iter(attrs=['pid', 'name', 'cmdline']) if './src/train.py' in p.info['cmdline']]
 	except Exception as e:
-		print(e)
 		pass
 
 	training_state.process.stdout.close()
@@ -1000,30 +1004,46 @@ 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)
 
-	res = whisper_model.transcribe(file)
-	segments = whisper_model.extract_text_and_timestamps( res )
+	elif args.whisper_backend == "lightmare/whispercpp":
+		res = whisper_model.transcribe(file)
+		segments = whisper_model.extract_text_and_timestamps( res )
 
-	result = {
-		'segments': []
-	}
-	for segment in segments:
-		reparsed = {
-			'start': segment[0] / 100.0,
-			'end': segment[1] / 100.0,
-			'text': segment[2],
+		result = {
+			'segments': []
 		}
-		result['segments'].append(reparsed)
+		for segment in segments:
+			reparsed = {
+				'start': segment[0] / 100.0,
+				'end': segment[1] / 100.0,
+				'text': segment[2],
+			}
+			result['segments'].append(reparsed)
 
-	return result
+		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)
 
-def prepare_dataset( files, outdir, language=None, progress=None ):
+		for i in range(len(result_aligned['segments'])):
+			del result_aligned['segments'][i]['word-segments']
+			del result_aligned['segments'][i]['char-segments']
+
+		result['segments'] = result_aligned['segments']
+
+		return result
+
+def prepare_dataset( files, outdir, language=None, skip_existings=False, progress=None ):
 	unload_tts()
 
 	global whisper_model
@@ -1034,9 +1054,30 @@ def prepare_dataset( files, outdir, language=None, progress=None ):
 
 	results = {}
 	transcription = []
+	files = sorted(files)
+
+	previous_list = []
+	parsed_list = []
+	if skip_existings and os.path.exists(f'{outdir}/train.txt'):
+		with open(f'{outdir}/train.txt', 'r', encoding="utf-8") as f:
+			parsed_list = f.readlines()
+
+		for line in parsed_list:
+			match = re.findall(r"^(.+?)_\d+\.wav$", line.split("|")[0])
+			print(match)
+			if match is None or len(match) == 0:
+				continue
+			
+			if match[0] not in previous_list:
+				previous_list.append(f'{match[0]}.wav')
 
 	for file in enumerate_progress(files, desc="Iterating through voice files", progress=progress):
 		basename = os.path.basename(file)
+
+		if basename in previous_list:
+			print(f"Skipping already parsed file: {basename}")
+			continue
+
 		result = whisper_transcribe(file, language=language)
 		results[basename] = result
 		print(f"Transcribed file: {file}, {len(result['segments'])} found.")
@@ -1063,9 +1104,14 @@ def prepare_dataset( files, outdir, language=None, progress=None ):
 			transcription.append(line)
 			with open(f'{outdir}/train.txt', 'a', encoding="utf-8") as f:
 				f.write(f'{line}\n')
+
+		do_gc()
 	
 	with open(f'{outdir}/whisper.json', 'w', encoding="utf-8") as f:
 		f.write(json.dumps(results, indent='\t'))
+
+	if len(parsed_list) > 0:
+		transcription = parsed_list + transcription
 	
 	joined = '\n'.join(transcription)
 	with open(f'{outdir}/train.txt', 'w', encoding="utf-8") as f:
@@ -1419,13 +1465,13 @@ def setup_args():
 		'prune-nonfinal-outputs': True,
 		'use-bigvgan-vocoder': True,
 		'concurrency-count': 2,
-		'autocalculate-voice-chunk-duration-size': 10,
+		'autocalculate-voice-chunk-duration-size': 0,
 		'output-sample-rate': 44100,
 		'output-volume': 1,
 		
 		'autoregressive-model': None,
+		'whisper-backend': 'openai/whisper',
 		'whisper-model': "base",
-		'whisper-cpp': False,
 
 		'training-default-halfp': False,
 		'training-default-bnb': True,
@@ -1463,8 +1509,8 @@ 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")
@@ -1498,7 +1544,7 @@ 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
@@ -1522,8 +1568,8 @@ 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
@@ -1554,8 +1600,8 @@ 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,
@@ -1760,6 +1806,12 @@ 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
@@ -1773,16 +1825,20 @@ 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:
-		import whisper
-		whisper_model = whisper.load_model(model_name)
+	elif args.whisper_backend == "m-bain/whisperx":
+		import whisperx
+		device = "cuda" if get_device_name() == "cuda" else "cpu"
+		whisper_model = whisperx.load_model(model_name, device)
 
 	print("Loaded Whisper model")
 
@@ -1794,4 +1850,4 @@ def unload_whisper():
 		whisper_model = None
 		print("Unloaded Whisper")
 
-	do_gc()
+	do_gc()	

src/webui.py

@@ -180,13 +180,13 @@ def read_generate_settings_proxy(file, saveAs='.temp'):
 
 	return (
 		gr.update(value=j, visible=j is not None),
-		gr.update(visible=j is not None),
 		gr.update(value=latents, visible=latents is not None),
-		None if j is None else j['voice']
+		None if j is None else j['voice'],
+		gr.update(visible=j is not None),
 	)
 
-def prepare_dataset_proxy( voice, language, progress=gr.Progress(track_tqdm=True) ):
-	return prepare_dataset( get_voices(load_latents=False)[voice], outdir=f"./training/{voice}/", language=language, progress=progress )
+def prepare_dataset_proxy( voice, language, skip_existings, progress=gr.Progress(track_tqdm=True) ):
+	return prepare_dataset( get_voices(load_latents=False)[voice], outdir=f"./training/{voice}/", language=language, skip_existings=skip_existings, progress=progress )
 
 def optimize_training_settings_proxy( *args, **kwargs ):
 	tup = optimize_training_settings(*args, **kwargs)
@@ -378,15 +378,15 @@ def setup_gradio():
 		with gr.Tab("Generate"):
 			with gr.Row():
 				with gr.Column():
-					text = gr.Textbox(lines=4, label="Prompt")
+					text = gr.Textbox(lines=4, label="Input Prompt")
 			with gr.Row():
 				with gr.Column():
 					delimiter = gr.Textbox(lines=1, label="Line Delimiter", placeholder="\\n")
 
-					emotion = gr.Radio( ["Happy", "Sad", "Angry", "Disgusted", "Arrogant", "Custom"], value="Custom", label="Emotion", type="value", interactive=True )
-					prompt = gr.Textbox(lines=1, label="Custom Emotion + Prompt (if selected)")
+					emotion = gr.Radio( ["Happy", "Sad", "Angry", "Disgusted", "Arrogant", "Custom", "None"], value="None", label="Emotion", type="value", interactive=True )
+					prompt = gr.Textbox(lines=1, label="Custom Emotion")
 					voice = gr.Dropdown(choices=voice_list_with_defaults, label="Voice", type="value", value=voice_list_with_defaults[0]) # it'd be very cash money if gradio was able to default to the first value in the list without this shit
-					mic_audio = gr.Audio( label="Microphone Source", source="microphone", type="filepath" )
+					mic_audio = gr.Audio( label="Microphone Source", source="microphone", type="filepath", visible=False )
 					voice_latents_chunks = gr.Slider(label="Voice Chunks", minimum=1, maximum=128, value=1, step=1)
 					with gr.Row():
 						refresh_voices = gr.Button(value="Refresh Voice List")
@@ -397,6 +397,11 @@ def setup_gradio():
 						inputs=voice,
 						outputs=voice_latents_chunks
 					)
+					voice.change(
+						fn=lambda value: gr.update(visible=value == "microphone"),
+						inputs=voice,
+						outputs=mic_audio,
+					)
 				with gr.Column():
 					candidates = gr.Slider(value=1, minimum=1, maximum=6, step=1, label="Candidates")
 					seed = gr.Number(value=0, precision=0, label="Seed")
@@ -406,16 +411,17 @@ def setup_gradio():
 					diffusion_iterations = gr.Slider(value=128, minimum=0, maximum=512, step=1, label="Iterations")
 
 					temperature = gr.Slider(value=0.2, minimum=0, maximum=1, step=0.1, label="Temperature")
-					breathing_room = gr.Slider(value=8, minimum=1, maximum=32, step=1, label="Pause Size")
-					diffusion_sampler = gr.Radio(
-						["P", "DDIM"], # + ["K_Euler_A", "DPM++2M"],
-						value="P", label="Diffusion Samplers", type="value" )
 					show_experimental_settings = gr.Checkbox(label="Show Experimental Settings")
 					reset_generation_settings_button = gr.Button(value="Reset to Default")
 				with gr.Column(visible=False) as col:
 					experimental_column = col
 
 					experimental_checkboxes = gr.CheckboxGroup(["Half Precision", "Conditioning-Free"], value=["Conditioning-Free"], label="Experimental Flags")
+					breathing_room = gr.Slider(value=8, minimum=1, maximum=32, step=1, label="Pause Size")
+					diffusion_sampler = gr.Radio(
+						["P", "DDIM"], # + ["K_Euler_A", "DPM++2M"],
+						value="DDIM", label="Diffusion Samplers", type="value"
+					)
 					cvvp_weight = gr.Slider(value=0, minimum=0, maximum=1, label="CVVP Weight")
 					top_p = gr.Slider(value=0.8, minimum=0, maximum=1, label="Top P")
 					diffusion_temperature = gr.Slider(value=1.0, minimum=0, maximum=1, label="Diffusion Temperature")
@@ -460,17 +466,20 @@ def setup_gradio():
 					audio_in = gr.Files(type="file", label="Audio Input", file_types=["audio"])
 					import_voice_name = gr.Textbox(label="Voice Name")
 					import_voice_button = gr.Button(value="Import Voice")
-				with gr.Column():
-					metadata_out = gr.JSON(label="Audio Metadata", visible=False)
-					copy_button = gr.Button(value="Copy Settings", visible=False)
-					latents_out = gr.File(type="binary", label="Voice Latents", visible=False)
+				with gr.Column(visible=False) as col:
+					utilities_metadata_column = col
+
+					metadata_out = gr.JSON(label="Audio Metadata")
+					copy_button = gr.Button(value="Copy Settings")
+					latents_out = gr.File(type="binary", label="Voice Latents")
 		with gr.Tab("Training"):
 			with gr.Tab("Prepare Dataset"):
 				with gr.Row():
 					with gr.Column():
 						dataset_settings = [
 							gr.Dropdown( choices=voice_list, label="Dataset Source", type="value", value=voice_list[0] if len(voice_list) > 0 else "" ),
-							gr.Textbox(label="Language", value="en")
+							gr.Textbox(label="Language", value="en"),
+							gr.Checkbox(label="Skip Already Transcribed", value=False)
 						]
 						prepare_dataset_button = gr.Button(value="Prepare")
 					with gr.Column():
@@ -581,10 +590,10 @@ 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")
@@ -662,9 +671,9 @@ def setup_gradio():
 			inputs=audio_in,
 			outputs=[
 				metadata_out,
-				copy_button,
 				latents_out,
-				import_voice_name
+				import_voice_name,
+				utilities_metadata_column,
 			]
 		)
 
@@ -697,9 +706,10 @@ def setup_gradio():
 			outputs=voice,
 		)
 		
-		prompt.change(fn=lambda value: gr.update(value="Custom"),
-			inputs=prompt,
-			outputs=emotion
+		emotion.change(
+			fn=lambda value: gr.update(visible=value == "Custom"),
+			inputs=emotion,
+			outputs=prompt
 		)
 		mic_audio.change(fn=lambda value: gr.update(value="microphone"),
 			inputs=mic_audio,
@@ -723,6 +733,7 @@ def setup_gradio():
 		submit_event = submit.click(run_generation,
 			inputs=input_settings,
 			outputs=[output_audio, source_sample, candidates_list, generation_results],
+			api_name="generate",
 		)
 
 

start.sh

@@ -1,4 +1,6 @@
 #!/bin/bash
+if [ ! -d "venv" ]; then ./setup-guided.sh; fi
+
 source ./venv/bin/activate
 python3 ./src/main.py "$@"
 deactivate

update-force.bat

@@ -6,10 +6,10 @@ python -m venv venv
 call .\venv\Scripts\activate.bat
 
 python -m pip install --upgrade pip
-python -m pip install -U -r .\dlas\requirements.txt
+python -m pip install -U -r .\requirements.txt
 python -m pip install -U -r .\tortoise-tts\requirements.txt
 python -m pip install -U -e .\tortoise-tts 
-python -m pip install -U -r .\requirements.txt
+python -m pip install -U -r .\dlas\requirements.txt
 
 pause
 deactivate

update-force.sh

@@ -9,9 +9,9 @@ python3 -m venv venv
 source ./venv/bin/activate
 
 python3 -m pip install --upgrade pip
-python3 -m pip install -r ./dlas/requirements.txt
+python3 -m pip install -r ./requirements.txt
 python3 -m pip install -r ./tortoise-tts/requirements.txt
 python3 -m pip install -e ./tortoise-tts 
-python3 -m pip install -r ./requirements.txt
+python3 -m pip install -r ./dlas/requirements.txt
 
 deactivate

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