Spaces:
Sleeping
Sleeping
File size: 11,693 Bytes
7694c84 |
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 |
from typing import List, Union, Optional, Literal
import text
import torch
import torch.nn as nn
from utils import get_basic_config
from vocoder import load_hifigan
from vocoder.hifigan.denoiser import Denoiser
from .fastpitch.model import FastPitch as _FastPitch
from ..diacritizers import load_vowelizer
_VOWELIZER_TYPE = Literal['shakkala', 'shakkelha']
def text_collate_fn(batch: List[torch.Tensor]):
"""
Args:
batch: List[text_ids]
Returns:
text_ids_pad
input_lens_sorted
reverse_ids
"""
input_lens_sorted, input_sort_ids = torch.sort(
torch.LongTensor([len(x) for x in batch]), descending=True)
max_input_len = input_lens_sorted[0]
text_ids_pad = torch.LongTensor(len(batch), max_input_len)
text_ids_pad.zero_()
for i in range(len(input_sort_ids)):
text_ids = batch[input_sort_ids[i]]
text_ids_pad[i, :text_ids.size(0)] = text_ids
return text_ids_pad, input_lens_sorted, input_sort_ids.argsort()
class FastPitch(_FastPitch):
def __init__(self,
checkpoint: str,
arabic_in: bool = True,
vowelizer: Optional[_VOWELIZER_TYPE] = None,
**kwargs):
from models.fastpitch import net_config
state_dicts = torch.load(checkpoint, map_location='cpu')
if 'config' in state_dicts:
net_config = state_dicts['config']
super().__init__(**net_config)
#self.n_eos = len(EOS_TOKENS)
self.arabic_in = arabic_in
#if checkpoint is not None:
self.load_state_dict(state_dicts['model'])
self.config = get_basic_config()
self.vowelizers = {}
if vowelizer is not None:
self.vowelizers[vowelizer] = load_vowelizer(vowelizer, self.config)
self.default_vowelizer = vowelizer
self.phon_to_id = None
if 'symbols' in state_dicts:
self.phon_to_id = {phon: i for i, phon in enumerate(state_dicts['symbols'])}
self.eval()
@property
def device(self):
return next(self.parameters()).device
def _vowelize(self, utterance: str, vowelizer: Optional[_VOWELIZER_TYPE] = None):
vowelizer = self.default_vowelizer if vowelizer is None else vowelizer
if vowelizer is not None:
if not vowelizer in self.vowelizers:
self.vowelizers[vowelizer] = load_vowelizer(vowelizer, self.config)
# print(f"loaded: {vowelizer}")
utterance_ar = text.buckwalter_to_arabic(utterance)
utterance = self.vowelizers[vowelizer].predict(utterance_ar)
return utterance
def _tokenize(self, utterance: str, vowelizer: Optional[_VOWELIZER_TYPE] = None):
utterance = self._vowelize(utterance=utterance, vowelizer=vowelizer)
if self.arabic_in:
return text.arabic_to_tokens(utterance, append_space=False)
return text.buckwalter_to_tokens(utterance, append_space=False)
@torch.inference_mode()
def ttmel_single(self,
utterance: str,
speed: float = 1,
speaker_id: int = 0,
vowelizer: Optional[_VOWELIZER_TYPE] = None
):
tokens = self._tokenize(utterance, vowelizer=vowelizer)
token_ids = text.tokens_to_ids(tokens, self.phon_to_id)
ids_batch = torch.LongTensor(token_ids).unsqueeze(0).to(self.device)
sid = torch.LongTensor([speaker_id]).to(self.device)
# Infer spectrogram
mel_spec, *_ = self.infer(ids_batch,
pace=speed,
speaker=speaker_id)
mel_spec = mel_spec[0]
return mel_spec # [F, T]
@torch.inference_mode()
def ttmel_batch(self,
batch: List[str],
speed: float = 1,
speaker_id: int = 0,
vowelizer: Optional[_VOWELIZER_TYPE] = None
):
batch_tokens = [
self._tokenize(line, vowelizer=vowelizer)
for line in batch
]
batch_ids = [torch.LongTensor(
text.tokens_to_ids(tokens, self.phon_to_id)
) for tokens in batch_tokens]
batch = text_collate_fn(batch_ids)
(
batch_ids_padded, batch_lens_sorted,
reverse_sort_ids
) = batch
batch_ids_padded = batch_ids_padded.to(self.device)
batch_lens_sorted = batch_lens_sorted.to(self.device)
batch_sids = batch_lens_sorted*0 + speaker_id
y_pred = self.infer(batch_ids_padded, pace=speed, speaker=speaker_id)
mel_outputs, mel_specgram_lengths, *_ = y_pred
mel_list = []
for i, id in enumerate(reverse_sort_ids):
mel = mel_outputs[id, :, :mel_specgram_lengths[id]]
mel_list.append(mel)
return mel_list
def ttmel(self,
text_input: Union[str, List[str]],
speed: float = 1,
speaker_id: int = 0,
batch_size: int = 1,
vowelizer: Optional[_VOWELIZER_TYPE] = None
):
# input: string
if isinstance(text_input, str):
return self.ttmel_single(text_input, speed=speed,
speaker_id=speaker_id,
vowelizer=vowelizer)
# input: list
assert isinstance(text_input, list)
batch = text_input
mel_list = []
if batch_size == 1:
for sample in batch:
mel = self.ttmel_single(sample, speed=speed,
speaker_id=speaker_id,
vowelizer=vowelizer)
mel_list.append(mel)
return mel_list
# infer one batch
if len(batch) <= batch_size:
return self.ttmel_batch(batch, speed=speed,
speaker_id=speaker_id,
vowelizer=vowelizer)
# batched inference
batches = [batch[k:k+batch_size]
for k in range(0, len(batch), batch_size)]
for batch in batches:
mels = self.ttmel_batch(batch, speed=speed,
speaker_id=speaker_id,
vowelizer=vowelizer)
mel_list += mels
return mel_list
class FastPitch2Wave(nn.Module):
def __init__(self,
model_sd_path: str,
vocoder_sd: Optional[str] = None,
vocoder_config: Optional[str] = None,
vowelizer: Optional[_VOWELIZER_TYPE] = None,
arabic_in: bool = True,
):
super().__init__()
# from models.fastpitch import net_config
state_dicts = torch.load(model_sd_path, map_location='cpu')
# if 'config' in state_dicts:
# net_config = state_dicts['config']
model = FastPitch(model_sd_path,
arabic_in=arabic_in,
vowelizer=vowelizer)
model.load_state_dict(state_dicts['model'], strict=False)
self.model = model
if vocoder_sd is None or vocoder_config is None:
config = get_basic_config()
vocoder_sd = config.vocoder_state_path
vocoder_config = config.vocoder_config_path
vocoder = load_hifigan(vocoder_sd, vocoder_config)
self.vocoder = vocoder
self.denoiser = Denoiser(vocoder)
self.eval()
@property
def device(self):
return next(self.parameters()).device
def forward(self, x):
return x
@torch.inference_mode()
def tts_single(self,
text_buckw: str,
speed: float = 1,
speaker_id: int = 0,
denoise: float = 0,
vowelizer: Optional[_VOWELIZER_TYPE] = None,
return_mel=False):
mel_spec = self.model.ttmel_single(text_buckw, speed,
speaker_id, vowelizer)
wave = self.vocoder(mel_spec)
if denoise > 0:
wave = self.denoiser(wave, denoise)
if return_mel:
return wave[0].cpu(), mel_spec
return wave[0].cpu()
@torch.inference_mode()
def tts_batch(self,
batch: List[str],
speed: float = 1,
speaker_id: int = 0,
denoise: float = 0,
vowelizer: Optional[_VOWELIZER_TYPE] = None,
return_mel: bool = False
):
mel_list = self.model.ttmel_batch(batch, speed,
speaker_id, vowelizer)
wav_list = []
for mel in mel_list:
wav_inferred = self.vocoder(mel)
if denoise > 0:
wav_inferred = self.denoiser(wav_inferred, denoise)
wav_list.append(wav_inferred[0].cpu())
if return_mel:
wav_list, mel_list
return wav_list
def tts(self,
text_input: Union[str, List[str]],
speed: float = 1.,
denoise: float = 0,
speaker_id: int = 0,
batch_size: int = 2,
vowelizer: Optional[_VOWELIZER_TYPE] = None,
return_mel: bool = False):
"""
Args:
text_input (str|List[str]): Input text.
speed (float): Speaking speed.
denoise (float): Hifi-GAN Denoiser strength.
speaker_id (int): Speaker Id.
batch_size (int): bacch size for inferrence.
vowelizer (None|str): options [None, `'shakkala'`, `'shakkelha'`].
return_mel (bool): Whether to return the mel spectrogram(s).
"""
# input: string
if isinstance(text_input, str):
return self.tts_single(text_input, speaker_id=speaker_id,
speed=speed, denoise=denoise,
vowelizer=vowelizer,
return_mel=return_mel)
# input: list
assert isinstance(text_input, list)
batch = text_input
wav_list = []
if batch_size == 1:
for sample in batch:
wav = self.tts_single(sample, speaker_id=speaker_id,
speed=speed, denoise=denoise,
vowelizer=vowelizer,
return_mel=return_mel)
wav_list.append(wav)
return wav_list
# infer one batch
if len(batch) <= batch_size:
return self.tts_batch(batch, speaker_id=speaker_id,
speed=speed, denoise=denoise,
vowelizer=vowelizer,
return_mel=return_mel)
# batched inference
batches = [batch[k:k+batch_size]
for k in range(0, len(batch), batch_size)]
for batch in batches:
wavs = self.tts_batch(batch, speaker_id=speaker_id,
speed=speed, denoise=denoise,
vowelizer=vowelizer,
return_mel=return_mel)
wav_list += wavs
return wav_list
|