added (yet another) experimental voice latent calculation mode (when chunk size is 0 and theres a dataset generated, itll leverage it by padding to a common size then computing them, should help avoid splitting mid-phoneme)

This commit is contained in:
mrq 2023-03-07 03:55:35 +00:00
parent 5063728bb0
commit 3899f9b4e3
2 changed files with 58 additions and 17 deletions

View File

@ -31,7 +31,7 @@ import pandas as pd
from datetime import datetime from datetime import datetime
from datetime import timedelta from datetime import timedelta
from tortoise.api import TextToSpeech, MODELS, get_model_path from tortoise.api import TextToSpeech, MODELS, get_model_path, pad_or_truncate
from tortoise.utils.audio import load_audio, load_voice, load_voices, get_voice_dir from tortoise.utils.audio import load_audio, load_voice, load_voices, get_voice_dir
from tortoise.utils.text import split_and_recombine_text from tortoise.utils.text import split_and_recombine_text
from tortoise.utils.device import get_device_name, set_device_name from tortoise.utils.device import get_device_name, set_device_name
@ -89,6 +89,8 @@ def generate(
if tts_loading: if tts_loading:
raise Exception("TTS is still initializing...") raise Exception("TTS is still initializing...")
load_tts() load_tts()
if hasattr(tts, "loading") and tts.loading:
raise Exception("TTS is still initializing...")
do_gc() do_gc()
@ -121,17 +123,8 @@ def generate(
voice_samples, conditioning_latents = load_voice(voice) voice_samples, conditioning_latents = load_voice(voice)
if voice_samples and len(voice_samples) > 0: if voice_samples and len(voice_samples) > 0:
conditioning_latents = compute_latents(voice=voice, voice_samples=voice_samples, voice_latents_chunks=voice_latents_chunks)
sample_voice = torch.cat(voice_samples, dim=-1).squeeze().cpu() sample_voice = torch.cat(voice_samples, dim=-1).squeeze().cpu()
conditioning_latents = tts.get_conditioning_latents(voice_samples, return_mels=not args.latents_lean_and_mean, progress=progress, slices=voice_latents_chunks, force_cpu=args.force_cpu_for_conditioning_latents)
if len(conditioning_latents) == 4:
conditioning_latents = (conditioning_latents[0], conditioning_latents[1], conditioning_latents[2], None)
if voice != "microphone":
if hasattr(tts, 'autoregressive_model_hash'):
torch.save(conditioning_latents, f'{get_voice_dir()}/{voice}/cond_latents_{tts.autoregressive_model_hash[:8]}.pth')
else:
torch.save(conditioning_latents, f'{get_voice_dir()}/{voice}/cond_latents.pth')
voice_samples = None voice_samples = None
else: else:
if conditioning_latents is not None: if conditioning_latents is not None:
@ -551,6 +544,10 @@ def update_baseline_for_latents_chunks( voice ):
if not os.path.isdir(path): if not os.path.isdir(path):
return 1 return 1
dataset_file = f'./training/{voice}/train.txt'
if os.path.exists(dataset_file):
return 0 # 0 will leverage using the LJspeech dataset for computing latents
files = os.listdir(path) files = os.listdir(path)
total = 0 total = 0
@ -565,11 +562,13 @@ def update_baseline_for_latents_chunks( voice ):
total_duration += duration total_duration += duration
total = total + 1 total = total + 1
# brain too fried to figure out a better way
if args.autocalculate_voice_chunk_duration_size == 0: if args.autocalculate_voice_chunk_duration_size == 0:
return int(total_duration / total) if total > 0 else 1 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 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)): def compute_latents(voice=None, voice_samples=None, voice_latents_chunks=0, progress=None):
global tts global tts
global args global args
@ -581,12 +580,42 @@ def compute_latents(voice, voice_latents_chunks, progress=gr.Progress(track_tqdm
raise Exception("TTS is still initializing...") raise Exception("TTS is still initializing...")
load_tts() load_tts()
voice_samples, conditioning_latents = load_voice(voice, load_latents=False) if hasattr(tts, "loading") and tts.loading:
raise Exception("TTS is still initializing...")
if voice:
load_from_dataset = voice_latents_chunks == 0
if load_from_dataset:
dataset_path = f'./training/{voice}/train.txt'
if not os.path.exists(dataset_path):
load_from_dataset = False
else:
with open(dataset_path, 'r', encoding="utf-8") as f:
lines = f.readlines()
print("Leveraging LJSpeech dataset for computing latents")
voice_samples = []
max_length = 0
for line in lines:
filename = f'./training/{voice}/{line.split("|")[0]}'
waveform = load_audio(filename, 22050)
max_length = max(max_length, waveform.shape[-1])
voice_samples.append(waveform)
for i in range(len(voice_samples)):
voice_samples[i] = pad_or_truncate(voice_samples[i], max_length)
voice_latents_chunks = len(voice_samples)
if not load_from_dataset:
voice_samples, _ = load_voice(voice, load_latents=False)
if voice_samples is None: if voice_samples is None:
return return
conditioning_latents = tts.get_conditioning_latents(voice_samples, return_mels=not args.latents_lean_and_mean, progress=progress, slices=voice_latents_chunks, force_cpu=args.force_cpu_for_conditioning_latents) conditioning_latents = tts.get_conditioning_latents(voice_samples, return_mels=not args.latents_lean_and_mean, slices=voice_latents_chunks, force_cpu=args.force_cpu_for_conditioning_latents, progress=progress)
if len(conditioning_latents) == 4: if len(conditioning_latents) == 4:
conditioning_latents = (conditioning_latents[0], conditioning_latents[1], conditioning_latents[2], None) conditioning_latents = (conditioning_latents[0], conditioning_latents[1], conditioning_latents[2], None)
@ -596,7 +625,7 @@ def compute_latents(voice, voice_latents_chunks, progress=gr.Progress(track_tqdm
else: else:
torch.save(conditioning_latents, f'{get_voice_dir()}/{voice}/cond_latents.pth') torch.save(conditioning_latents, f'{get_voice_dir()}/{voice}/cond_latents.pth')
return voice return conditioning_latents
# superfluous, but it cleans up some things # superfluous, but it cleans up some things
class TrainingState(): class TrainingState():
@ -1847,6 +1876,10 @@ def update_autoregressive_model(autoregressive_model_path):
if tts_loading: if tts_loading:
raise Exception("TTS is still initializing...") raise Exception("TTS is still initializing...")
return return
if hasattr(tts, "loading") and tts.loading:
raise Exception("TTS is still initializing...")
print(f"Loading model: {autoregressive_model_path}") print(f"Loading model: {autoregressive_model_path}")
tts.load_autoregressive_model(autoregressive_model_path) tts.load_autoregressive_model(autoregressive_model_path)
@ -1867,6 +1900,9 @@ def update_vocoder_model(vocoder_model):
raise Exception("TTS is still initializing...") raise Exception("TTS is still initializing...")
return return
if hasattr(tts, "loading") and tts.loading:
raise Exception("TTS is still initializing...")
print(f"Loading model: {vocoder_model}") print(f"Loading model: {vocoder_model}")
tts.load_vocoder_model(vocoder_model) tts.load_vocoder_model(vocoder_model)
print(f"Loaded model: {tts.vocoder_model}") print(f"Loaded model: {tts.vocoder_model}")

View File

@ -163,6 +163,11 @@ def history_view_results( voice ):
gr.Dropdown.update(choices=sorted(files)) gr.Dropdown.update(choices=sorted(files))
) )
def compute_latents_proxy(voice, voice_latents_chunks, progress=gr.Progress(track_tqdm=True)):
compute_latents( voice=voice, voice_latents_chunks=voice_latents_chunks, progress=progress )
return voice
def import_voices_proxy(files, name, progress=gr.Progress(track_tqdm=True)): def import_voices_proxy(files, name, progress=gr.Progress(track_tqdm=True)):
import_voices(files, name, progress) import_voices(files, name, progress)
return gr.update() return gr.update()
@ -387,7 +392,7 @@ def setup_gradio():
prompt = gr.Textbox(lines=1, label="Custom Emotion") 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 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", visible=False ) 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) voice_latents_chunks = gr.Number(label="Voice Chunks", precision=0, value=0)
with gr.Row(): with gr.Row():
refresh_voices = gr.Button(value="Refresh Voice List") refresh_voices = gr.Button(value="Refresh Voice List")
recompute_voice_latents = gr.Button(value="(Re)Compute Voice Latents") recompute_voice_latents = gr.Button(value="(Re)Compute Voice Latents")
@ -704,7 +709,7 @@ def setup_gradio():
], ],
) )
recompute_voice_latents.click(compute_latents, recompute_voice_latents.click(compute_latents_proxy,
inputs=[ inputs=[
voice, voice,
voice_latents_chunks, voice_latents_chunks,