File size: 11,606 Bytes
33e4bc7 73e9929 f37375b 33e4bc7 |
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 |
import argparse
import os
import random
from urllib import request
import torch
import torch.nn.functional as F
import torchaudio
import progressbar
import ocotillo
from models.diffusion_decoder import DiffusionTts
from models.autoregressive import UnifiedVoice
from tqdm import tqdm
from models.arch_util import TorchMelSpectrogram
from models.new_autoregressive import AutoregressiveCodegen
from models.text_voice_clip import VoiceCLIP
from models.vocoder import UnivNetGenerator
from utils.audio import load_audio, wav_to_univnet_mel, denormalize_tacotron_mel
from utils.diffusion import SpacedDiffusion, space_timesteps, get_named_beta_schedule
from utils.tokenizer import VoiceBpeTokenizer, lev_distance
pbar = None
def download_models():
MODELS = {
'clip.pth': 'https://huggingface.co/jbetker/tortoise-tts-clip/resolve/main/pytorch-model.bin',
'diffusion.pth': 'https://huggingface.co/jbetker/tortoise-tts-diffusion-v1/resolve/main/pytorch-model.bin',
'autoregressive.pth': 'https://huggingface.co/jbetker/tortoise-tts-autoregressive/resolve/main/pytorch-model.bin'
}
os.makedirs('.models', exist_ok=True)
def show_progress(block_num, block_size, total_size):
global pbar
if pbar is None:
pbar = progressbar.ProgressBar(maxval=total_size)
pbar.start()
downloaded = block_num * block_size
if downloaded < total_size:
pbar.update(downloaded)
else:
pbar.finish()
pbar = None
for model_name, url in MODELS.items():
if os.path.exists(f'.models/{model_name}'):
continue
print(f'Downloading {model_name} from {url}...')
request.urlretrieve(url, f'.models/{model_name}', show_progress)
print('Done.')
def pad_or_truncate(t, length):
if t.shape[-1] == length:
return t
elif t.shape[-1] < length:
return F.pad(t, (0, length-t.shape[-1]))
else:
return t[..., :length]
def load_discrete_vocoder_diffuser(trained_diffusion_steps=4000, desired_diffusion_steps=200, cond_free=True, cond_free_k=1):
"""
Helper function to load a GaussianDiffusion instance configured for use as a vocoder.
"""
return SpacedDiffusion(use_timesteps=space_timesteps(trained_diffusion_steps, [desired_diffusion_steps]), model_mean_type='epsilon',
model_var_type='learned_range', loss_type='mse', betas=get_named_beta_schedule('linear', trained_diffusion_steps),
conditioning_free=cond_free, conditioning_free_k=cond_free_k)
def load_conditioning(clip, cond_length=132300):
gap = clip.shape[-1] - cond_length
if gap < 0:
clip = F.pad(clip, pad=(0, abs(gap)))
elif gap > 0:
rand_start = random.randint(0, gap)
clip = clip[:, rand_start:rand_start + cond_length]
mel_clip = TorchMelSpectrogram()(clip.unsqueeze(0)).squeeze(0)
return mel_clip.unsqueeze(0).cuda()
def fix_autoregressive_output(codes, stop_token):
"""
This function performs some padding on coded audio that fixes a mismatch issue between what the diffusion model was
trained on and what the autoregressive code generator creates (which has no padding or end).
This is highly specific to the DVAE being used, so this particular coding will not necessarily work if used with
a different DVAE. This can be inferred by feeding a audio clip padded with lots of zeros on the end through the DVAE
and copying out the last few codes.
Failing to do this padding will produce speech with a harsh end that sounds like "BLAH" or similar.
"""
# Strip off the autoregressive stop token and add padding.
stop_token_indices = (codes == stop_token).nonzero()
if len(stop_token_indices) == 0:
print("No stop tokens found, enjoy that output of yours!")
return codes
else:
codes[stop_token_indices] = 83
stm = stop_token_indices.min().item()
codes[stm:] = 83
if stm - 3 < codes.shape[0]:
codes[-3] = 45
codes[-2] = 45
codes[-1] = 248
return codes
def do_spectrogram_diffusion(diffusion_model, diffuser, mel_codes, conditioning_samples, temperature=1):
"""
Uses the specified diffusion model to convert discrete codes into a spectrogram.
"""
with torch.no_grad():
cond_mels = []
for sample in conditioning_samples:
sample = pad_or_truncate(sample, 102400)
cond_mel = wav_to_univnet_mel(sample.to(mel_codes.device), do_normalization=False)
cond_mels.append(cond_mel)
cond_mels = torch.stack(cond_mels, dim=1)
output_seq_len = mel_codes.shape[-1]*4*24000//22050 # This diffusion model converts from 22kHz spectrogram codes to a 24kHz spectrogram signal.
output_shape = (mel_codes.shape[0], 100, output_seq_len)
precomputed_embeddings = diffusion_model.timestep_independent(mel_codes, cond_mels, output_seq_len, False)
noise = torch.randn(output_shape, device=mel_codes.device) * temperature
mel = diffuser.p_sample_loop(diffusion_model, output_shape, noise=noise,
model_kwargs={'precomputed_aligned_embeddings': precomputed_embeddings})
return denormalize_tacotron_mel(mel)[:,:,:output_seq_len]
class TextToSpeech:
def __init__(self, autoregressive_batch_size=32):
self.autoregressive_batch_size = autoregressive_batch_size
self.tokenizer = VoiceBpeTokenizer()
download_models()
self.autoregressive = AutoregressiveCodegen(1024, 16).cpu().eval()
self.autoregressive.load_state_dict(torch.load('X:\\dlas\\experiments\\train_autoregressive_codegen\\models\\17000_codegen_ema.pth'))
self.clip = VoiceCLIP(dim_text=512, dim_speech=512, dim_latent=512, num_text_tokens=256, text_enc_depth=12,
text_seq_len=350, text_heads=8,
num_speech_tokens=8192, speech_enc_depth=12, speech_heads=8, speech_seq_len=430,
use_xformers=True).cpu().eval()
self.clip.load_state_dict(torch.load('.models/clip.pth'))
self.diffusion = DiffusionTts(model_channels=1024, num_layers=10, in_channels=100, out_channels=200,
in_latent_channels=1024, in_tokens=8193, dropout=0, use_fp16=False, num_heads=16,
layer_drop=0, unconditioned_percentage=0).cpu().eval()
self.diffusion.load_state_dict(torch.load('.models/diffusion.pth'))
self.vocoder = UnivNetGenerator().cpu()
self.vocoder.load_state_dict(torch.load('.models/vocoder.pth')['model_g'])
self.vocoder.eval(inference=True)
def tts(self, text, voice_samples, k=1,
# autoregressive generation parameters follow
num_autoregressive_samples=512, temperature=.5, length_penalty=2, repetition_penalty=2.0, top_p=.5,
typical_sampling=False, typical_mass=.9,
# diffusion generation parameters follow
diffusion_iterations=100, cond_free=True, cond_free_k=2, diffusion_temperature=.7,):
text = torch.IntTensor(self.tokenizer.encode(text)).unsqueeze(0).cuda()
text = F.pad(text, (0, 1)) # This may not be necessary.
conds = []
if not isinstance(voice_samples, list):
voice_samples = [voice_samples]
for vs in voice_samples:
conds.append(load_conditioning(vs))
conds = torch.stack(conds, dim=1)
diffuser = load_discrete_vocoder_diffuser(desired_diffusion_steps=diffusion_iterations, cond_free=cond_free, cond_free_k=cond_free_k)
with torch.no_grad():
samples = []
num_batches = num_autoregressive_samples // self.autoregressive_batch_size
stop_mel_token = self.autoregressive.STOP_TOKEN
self.autoregressive = self.autoregressive.cuda()
for _ in tqdm(range(num_batches)):
codes = self.autoregressive.generate(conds, text,
do_sample=True,
top_p=top_p,
temperature=temperature,
num_return_sequences=self.autoregressive_batch_size,
length_penalty=length_penalty,
repetition_penalty=repetition_penalty,
typical_sampling=typical_sampling,
typical_mass=typical_mass)
padding_needed = 250 - codes.shape[1]
codes = F.pad(codes, (0, padding_needed), value=stop_mel_token)
samples.append(codes)
#self.autoregressive = self.autoregressive.cpu()
clip_results = []
self.clip = self.clip.cuda()
for batch in samples:
for i in range(batch.shape[0]):
batch[i] = fix_autoregressive_output(batch[i], stop_mel_token)
bad_toks = batch >= 8192
batch = batch * bad_toks.logical_not()
clip_results.append(self.clip(text.repeat(batch.shape[0], 1), batch, return_loss=False))
clip_results = torch.cat(clip_results, dim=0)
samples = torch.cat(samples, dim=0)
best_results = samples[torch.topk(clip_results, k=k).indices]
self.clip = self.clip.cpu()
del samples
print("Performing vocoding..")
wav_candidates = []
self.diffusion = self.diffusion.cuda()
self.vocoder = self.vocoder.cuda()
for b in range(best_results.shape[0]):
code = best_results[b].unsqueeze(0)
mel = do_spectrogram_diffusion(self.diffusion, diffuser, code, voice_samples, temperature=diffusion_temperature)
wav = self.vocoder.inference(mel)
wav_candidates.append(wav.cpu())
self.diffusion = self.diffusion.cpu()
self.vocoder = self.vocoder.cpu()
if len(wav_candidates) > 1:
return wav_candidates
return wav_candidates[0]
def refine_for_intellibility(self, wav_candidates, corresponding_codes, output_path):
"""
Further refine the remaining candidates using a ASR model to pick out the ones that are the most understandable.
TODO: finish this function
:param wav_candidates:
:return:
"""
transcriber = ocotillo.Transcriber(on_cuda=True)
transcriptions = transcriber.transcribe_batch(torch.cat(wav_candidates, dim=0).squeeze(1), 24000)
best = 99999999
for i, transcription in enumerate(transcriptions):
dist = lev_distance(transcription, args.text.lower())
if dist < best:
best = dist
best_codes = corresponding_codes[i].unsqueeze(0)
best_wav = wav_candidates[i]
del transcriber
torchaudio.save(os.path.join(output_path, f'{voice}_poor.wav'), best_wav.squeeze(0).cpu(), 24000)
# Perform diffusion again with the high-quality diffuser.
mel = do_spectrogram_diffusion(diffusion, final_diffuser, best_codes, cond_diffusion, mean=False)
wav = vocoder.inference(mel)
torchaudio.save(os.path.join(args.output_path, f'{voice}.wav'), wav.squeeze(0).cpu(), 24000) |