Dionyssos commited on
Commit
ab2d9a4
·
1 Parent(s): 1c1f53d

https://github.com/audeering/shift

Browse files
Files changed (1) hide show
  1. tts.py +847 -0
tts.py ADDED
@@ -0,0 +1,847 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import nltk
3
+ nltk.download('punkt', download_dir='./') # COMMENT IF DOWNLOADED
4
+ nltk.download('punkt_tab', download_dir='./') # COMMENT IF DOWNLOADED
5
+ nltk.data.path.append('.')
6
+ import librosa
7
+ import audiofile
8
+ import torch.nn.functional as F
9
+ import math
10
+ import numpy as np
11
+ import torch.nn as nn
12
+ import string
13
+ import textwrap
14
+ import phonemizer
15
+ from espeak_util import set_espeak_library
16
+ from transformers import AlbertConfig, AlbertModel
17
+ from huggingface_hub import hf_hub_download
18
+ from nltk.tokenize import word_tokenize
19
+ from torch.nn import Conv1d, ConvTranspose1d
20
+ from torch.nn.utils.parametrizations import weight_norm
21
+ from torch.nn.utils import spectral_norm
22
+
23
+ _pad = "$"
24
+ _punctuation = ';:,.!?¡¿—…"«»“” '
25
+ _letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
26
+ _letters_ipa = "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
27
+ MAX_PHONEMES = 424 # For OOM is the max length of single (non-split) sentence for StyleTTS2 inference
28
+
29
+ symbols = [_pad] + list(_punctuation) + list(_letters) + list(_letters_ipa)
30
+
31
+ dicts = {}
32
+ for i in range(len((symbols))):
33
+ dicts[symbols[i]] = i
34
+
35
+
36
+ class TextCleaner:
37
+ def __init__(self, dummy=None):
38
+ self.word_index_dictionary = dicts
39
+ print(len(dicts))
40
+
41
+ def __call__(self, text):
42
+ indexes = []
43
+ for char in text:
44
+ try:
45
+ indexes.append(self.word_index_dictionary[char])
46
+ except KeyError:
47
+ # `=NONVOCAL == \x00\x01\x02\x03\x04\x05\x06\x07\x08\t\n\x0b\x0c\r\x0e\x0f\x10\x11\x12\x13\x14\x15\x16\x17\x18\x19\x1a\x1b\x1c\x1d\x1e\x1f !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~\x7f
48
+ # print(f'NonVOCAL {char}', end='\r')
49
+ pass
50
+ return indexes
51
+
52
+ set_espeak_library()
53
+
54
+ textclenaer = TextCleaner()
55
+
56
+ global_phonemizer = phonemizer.backend.EspeakBackend(language="en-us", preserve_punctuation=True, with_stress=True)
57
+
58
+ def _del_prefix(d):
59
+ # del ".module"
60
+ out = {}
61
+ for k, v in d.items():
62
+ out[k[7:]] = v
63
+ return out
64
+
65
+
66
+
67
+
68
+ class StyleTTS2(nn.Module):
69
+
70
+ def __init__(self):
71
+ super().__init__()
72
+ albert_base_configuration = AlbertConfig(vocab_size=178,
73
+ hidden_size=768,
74
+ num_attention_heads=12,
75
+ intermediate_size=2048,
76
+ max_position_embeddings=512,
77
+ num_hidden_layers=12,
78
+ dropout=0.1)
79
+ self.bert = AlbertModel(albert_base_configuration)
80
+ state_dict = torch.load(hf_hub_download(repo_id='dkounadis/artificial-styletts2',
81
+ filename='Utils/PLBERT/step_1000000.pth'),
82
+ map_location='cpu')['net']
83
+ new_state_dict = {}
84
+ for k, v in state_dict.items():
85
+ name = k[7:] # remove `module.`
86
+ if name.startswith('encoder.'):
87
+ name = name[8:] # remove `encoder.`
88
+ new_state_dict[name] = v
89
+ del new_state_dict["embeddings.position_ids"]
90
+ self.bert.load_state_dict(new_state_dict, strict=True)
91
+ self.decoder = Decoder(dim_in=512,
92
+ style_dim=128,
93
+ dim_out=80, # n_mels
94
+ resblock_kernel_sizes=[3, 7, 11],
95
+ upsample_rates=[10, 5, 3, 2],
96
+ upsample_initial_channel=512,
97
+ resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
98
+ upsample_kernel_sizes=[20, 10, 6, 4])
99
+ self.text_encoder = TextEncoder(channels=512,
100
+ kernel_size=5,
101
+ depth=3, # args['model_params']['n_layer'],
102
+ n_symbols=178, # args['model_params']['n_token']
103
+ )
104
+ self.predictor = ProsodyPredictor(style_dim=128,
105
+ d_hid=512,
106
+ nlayers=3, # OFFICIAL config.nlayers=5;
107
+ max_dur=50)
108
+ self.style_encoder = StyleEncoder()
109
+ self.predictor_encoder = StyleEncoder()
110
+ self.bert_encoder = torch.nn.Linear(self.bert.config.hidden_size, 512)
111
+ self.mel_spec = MelSpec()
112
+ params = torch.load(hf_hub_download(repo_id='yl4579/StyleTTS2-LibriTTS',
113
+ filename='Models/LibriTTS/epochs_2nd_00020.pth'),
114
+ map_location='cpu')['net']
115
+ self.bert.load_state_dict(_del_prefix(params['bert']), strict=True)
116
+ self.bert_encoder.load_state_dict(_del_prefix(params['bert_encoder']), strict=True)
117
+ self.predictor.load_state_dict(_del_prefix(params['predictor']), strict=True)
118
+ self.decoder.load_state_dict(_del_prefix(params['decoder']), strict=True)
119
+ self.text_encoder.load_state_dict(_del_prefix(params['text_encoder']), strict=True)
120
+ self.predictor_encoder.load_state_dict(_del_prefix(params['predictor_encoder']), strict=True)
121
+ self.style_encoder.load_state_dict(_del_prefix(params['style_encoder']), strict=True)
122
+
123
+ # FOR LSTM
124
+ for n, p in self.named_parameters():
125
+ p.requires_grad = False
126
+ self.eval()
127
+
128
+
129
+ def device(self):
130
+ return self.style_encoder.unshared.weight.device
131
+
132
+ def compute_style(self, wav_file=None):
133
+
134
+ x, sr = librosa.load(wav_file, sr=24000)
135
+ x, _ = librosa.effects.trim(x, top_db=30)
136
+ if sr != 24000:
137
+ x = librosa.resample(x, sr, 24000)
138
+ # LOGMEL - Has 16KHz default basisc - Called on 24KHz .wav
139
+ x = torch.from_numpy(x[None, :]).to(device=self.device(),
140
+ dtype=torch.float)
141
+ mel_tensor = (torch.log(1e-5 + self.mel_spec(x)) + 4) / 4
142
+ #mel_tensor = preprocess(audio).to(device)
143
+ ref_s = self.style_encoder(mel_tensor)
144
+ ref_p = self.predictor_encoder(mel_tensor) # [bs, 11, 1, 128]
145
+ s = torch.cat([ref_s, ref_p], dim=3) # [bs, 11, 1, 256]
146
+ s = s[:, :, 0, :].transpose(1, 2) # [1, 128, 11]
147
+ return s # [1, 128, 11]
148
+
149
+ def inference(self,
150
+ text,
151
+ ref_s=None):
152
+ '''text may become too long when phonemized'''
153
+
154
+ if isinstance(ref_s, str):
155
+ ref_s = self.compute_style(ref_s)
156
+ else:
157
+ pass # assume ref_s = precomputed style vector
158
+
159
+
160
+ # text = transliterate_number(text, lang='en').strip()
161
+ # as we are in english transliteration is already done by the text cleaner?
162
+ # somehow we have phonemes in text that try to be rephonemized
163
+ # The ds txt should be only ascii
164
+
165
+
166
+ if isinstance(text, str):
167
+
168
+ _translator = str.maketrans('', '', string.punctuation)
169
+
170
+ text = [sub_sent.translate(_translator) + '.' for sub_sent in textwrap.wrap(text, 74)]
171
+
172
+ # # text = nltk.sent_tokenize(text)
173
+ # # text = [i for sent in sentences for i in textwrap.wrap(sent, width=120)]
174
+
175
+
176
+ # # text = textwrap.wrap(text, width=MAX_PHONEMES) # phonemes thus sent_tokenize() can't split them in sentences
177
+
178
+
179
+ device = ref_s.device
180
+ total = []
181
+ for _t in text:
182
+
183
+ _t = global_phonemizer.phonemize([_t])
184
+ _t = word_tokenize(_t[0])
185
+ _t = ' '.join(_t)
186
+
187
+ tokens = textclenaer(_t)[:MAX_PHONEMES] + [4] # textclenaer('.;?!') = [4,1,6,5] # append . punctuation to assure proper sound termination (pulse Issue)
188
+
189
+ # After filter we should assure is terminating as a sentence
190
+ # print(len(_t), len(tokens), 'Msi')#, textclenaer('.;?!'))
191
+ # ================================= Delete Phonemes If len(phonemes) > len(text) === OOM during training
192
+ tokens.insert(0, 0)
193
+ tokens = torch.LongTensor(tokens).to(device).unsqueeze(0)
194
+ with torch.no_grad():
195
+ hidden_states = self.text_encoder(tokens)
196
+ bert_dur = self.bert(tokens, attention_mask=torch.ones_like(tokens)
197
+ ).last_hidden_state
198
+ d_en = self.bert_encoder(bert_dur).transpose(-1, -2)
199
+ aln_trg, F0_pred, N_pred = self.predictor(d_en=d_en, s=ref_s[:, 128:, :])
200
+ asr = torch.bmm(aln_trg, hidden_states)
201
+ asr = asr.transpose(1, 2)
202
+ asr_new = torch.zeros_like(asr)
203
+ asr_new[:, :, 0] = asr[:, :, 0]
204
+ asr_new[:, :, 1:] = asr[:, :, 0:-1]
205
+ asr = asr_new
206
+ x = self.decoder(asr=asr,
207
+ F0_curve=F0_pred,
208
+ N=N_pred,
209
+ s=ref_s[:, :128, :]) # different part of ref_s
210
+ # print(x.shape, 'TTS TTS TTS TTS')
211
+ if x.shape[2] < 100:
212
+ x = torch.zeros(1, 1, 1000, device=self.device()) # silence if this sentence was empty
213
+
214
+ # NORMALIS / Crop Scratch at end (The endingscratch sound is not solved even with nltk.sentence split & punctuation)
215
+ x = x[..., 40:-4000]
216
+ # x /= x.abs().max() + 1e-7 # preserve as torch
217
+ # return x
218
+ if x.shape[2] == 0:
219
+ # nohing to vocode
220
+ x = torch.zeros(1, 1, 1000, device=self.device())
221
+ total.append(x)
222
+
223
+ # --
224
+ total = 1.94 * torch.cat(total, 2) # 1.94 * Perhaps exceeding -1,1 affects MIMI encode
225
+ total /= 1.02 * total.abs().max() + 1e-7
226
+ # --
227
+ return total
228
+
229
+
230
+
231
+
232
+ def get_padding(kernel_size, dilation=1):
233
+ return int((kernel_size*dilation - dilation)/2)
234
+
235
+
236
+ def _tile(x,
237
+ length=None):
238
+ x = x.repeat(1, 1, int(length / x.shape[2]) + 1)[:, :, :length]
239
+ return x
240
+
241
+
242
+ class AdaIN1d(nn.Module):
243
+
244
+ # used by HiFiGan & ProsodyPredictor
245
+
246
+ def __init__(self, style_dim, num_features):
247
+ super().__init__()
248
+ self.norm = nn.InstanceNorm1d(num_features, affine=False)
249
+ self.fc = nn.Linear(style_dim, num_features*2)
250
+
251
+ def forward(self, x, s):
252
+
253
+ # x = torch.Size([1, 512, 248]) same as output
254
+ # s = torch.Size([1, 7, 1, 128])
255
+
256
+ s = self.fc(s.transpose(1, 2)).transpose(1, 2)
257
+
258
+ s = _tile(s, length=x.shape[2])
259
+
260
+ gamma, beta = torch.chunk(s, chunks=2, dim=1)
261
+ return (1+gamma) * self.norm(x) + beta
262
+
263
+
264
+ class AdaINResBlock1(torch.nn.Module):
265
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5), style_dim=64):
266
+ super(AdaINResBlock1, self).__init__()
267
+ self.convs1 = nn.ModuleList([
268
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
269
+ padding=get_padding(kernel_size, dilation[0]))),
270
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
271
+ padding=get_padding(kernel_size, dilation[1]))),
272
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
273
+ padding=get_padding(kernel_size, dilation[2])))
274
+ ])
275
+ # self.convs1.apply(init_weights)
276
+
277
+ self.convs2 = nn.ModuleList([
278
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
279
+ padding=get_padding(kernel_size, 1))),
280
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
281
+ padding=get_padding(kernel_size, 1))),
282
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
283
+ padding=get_padding(kernel_size, 1)))
284
+ ])
285
+ # self.convs2.apply(init_weights)
286
+
287
+ self.adain1 = nn.ModuleList([
288
+ AdaIN1d(style_dim, channels),
289
+ AdaIN1d(style_dim, channels),
290
+ AdaIN1d(style_dim, channels),
291
+ ])
292
+
293
+ self.adain2 = nn.ModuleList([
294
+ AdaIN1d(style_dim, channels),
295
+ AdaIN1d(style_dim, channels),
296
+ AdaIN1d(style_dim, channels),
297
+ ])
298
+
299
+ self.alpha1 = nn.ParameterList(
300
+ [nn.Parameter(torch.ones(1, channels, 1)) for i in range(len(self.convs1))])
301
+ self.alpha2 = nn.ParameterList(
302
+ [nn.Parameter(torch.ones(1, channels, 1)) for i in range(len(self.convs2))])
303
+
304
+ def forward(self, x, s):
305
+ for c1, c2, n1, n2, a1, a2 in zip(self.convs1, self.convs2, self.adain1, self.adain2, self.alpha1, self.alpha2):
306
+ xt = n1(x, s) # THIS IS ADAIN - EXPECTS conv1d dims
307
+ xt = xt + (1 / a1) * (torch.sin(a1 * xt) ** 2) # Snake1D
308
+ xt = c1(xt)
309
+ xt = n2(xt, s) # THIS IS ADAIN - EXPECTS conv1d dims
310
+ xt = xt + (1 / a2) * (torch.sin(a2 * xt) ** 2) # Snake1D
311
+ xt = c2(xt)
312
+ x = xt + x
313
+ return x
314
+
315
+
316
+ class SourceModuleHnNSF(torch.nn.Module):
317
+
318
+ def __init__(self):
319
+
320
+ super().__init__()
321
+ self.harmonic_num = 8
322
+ self.l_linear = torch.nn.Linear(self.harmonic_num + 1, 1)
323
+ self.upsample_scale = 300
324
+
325
+
326
+ def forward(self, x):
327
+ # --
328
+ x = torch.multiply(x, torch.FloatTensor(
329
+ [[range(1, self.harmonic_num + 2)]]).to(x.device)) # [1, 145200, 9]
330
+
331
+ # modulo of negative f0_values => -21 % 10 = 9 as -3*10 + 9 = 21 NOTICE THAT f0_values IS SIGNED
332
+ rad_values = x / 25647 #).clamp(0, 1)
333
+ # rad_values = torch.where(torch.logical_or(rad_values < 0, rad_values > 1), 0.5, rad_values)
334
+ rad_values = rad_values % 1 # % of neg values
335
+ rad_values = F.interpolate(rad_values.transpose(1, 2),
336
+ scale_factor=1/self.upsample_scale,
337
+ mode='linear').transpose(1, 2)
338
+
339
+ # 1.89 sounds also nice has woofer at punctuation
340
+ phase = torch.cumsum(rad_values, dim=1) * 1.84 * np.pi
341
+ phase = F.interpolate(phase.transpose(1, 2) * self.upsample_scale,
342
+ scale_factor=self.upsample_scale, mode='linear').transpose(1, 2)
343
+ x = .009 * phase.sin()
344
+ # --
345
+ x = self.l_linear(x).tanh()
346
+ return x
347
+
348
+
349
+ class Generator(torch.nn.Module):
350
+ def __init__(self,
351
+ style_dim,
352
+ resblock_kernel_sizes,
353
+ upsample_rates,
354
+ upsample_initial_channel,
355
+ resblock_dilation_sizes,
356
+ upsample_kernel_sizes):
357
+ super(Generator, self).__init__()
358
+ self.num_kernels = len(resblock_kernel_sizes)
359
+ self.num_upsamples = len(upsample_rates)
360
+ self.m_source = SourceModuleHnNSF()
361
+ self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates))
362
+ self.noise_convs = nn.ModuleList()
363
+ self.ups = nn.ModuleList()
364
+ self.noise_res = nn.ModuleList()
365
+
366
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
367
+ c_cur = upsample_initial_channel // (2 ** (i + 1))
368
+
369
+ self.ups.append(weight_norm(ConvTranspose1d(upsample_initial_channel//(2**i),
370
+ upsample_initial_channel//(
371
+ 2**(i+1)),
372
+ k, u, padding=(u//2 + u % 2), output_padding=u % 2)))
373
+
374
+ if i + 1 < len(upsample_rates):
375
+ stride_f0 = np.prod(upsample_rates[i + 1:])
376
+ self.noise_convs.append(Conv1d(
377
+ 1, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=(stride_f0+1) // 2))
378
+ self.noise_res.append(AdaINResBlock1(
379
+ c_cur, 7, [1, 3, 5], style_dim))
380
+ else:
381
+ self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
382
+ self.noise_res.append(AdaINResBlock1(
383
+ c_cur, 11, [1, 3, 5], style_dim))
384
+
385
+ self.resblocks = nn.ModuleList()
386
+
387
+ self.alphas = nn.ParameterList()
388
+ self.alphas.append(nn.Parameter(
389
+ torch.ones(1, upsample_initial_channel, 1)))
390
+
391
+ for i in range(len(self.ups)):
392
+ ch = upsample_initial_channel//(2**(i+1))
393
+ self.alphas.append(nn.Parameter(torch.ones(1, ch, 1)))
394
+
395
+ for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
396
+ self.resblocks.append(AdaINResBlock1(ch, k, d, style_dim))
397
+
398
+ self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
399
+
400
+ def forward(self, x, s, f0):
401
+
402
+ # x.shape=torch.Size([1, 512, 484]) s.shape=torch.Size([1, 1, 1, 128]) f0.shape=torch.Size([1, 484]) GENERAT 249
403
+ f0 = self.f0_upsamp(f0).transpose(1, 2)
404
+
405
+ # x.shape=torch.Size([1, 512, 484]) s.shape=torch.Size([1, 1, 1, 128]) f0.shape=torch.Size([1, 145200, 1]) GENERAT 253
406
+
407
+ # [1, 145400, 1] f0 enters already upsampled to full wav 24kHz length
408
+ har_source = self.m_source(f0)
409
+
410
+ har_source = har_source.transpose(1, 2)
411
+
412
+ for i in range(self.num_upsamples):
413
+
414
+ x = x + (1 / self.alphas[i]) * (torch.sin(self.alphas[i] * x) ** 2)
415
+ x_source = self.noise_convs[i](har_source)
416
+ x_source = self.noise_res[i](x_source, s)
417
+
418
+ x = self.ups[i](x)
419
+
420
+ x = x + x_source
421
+
422
+ xs = None
423
+ for j in range(self.num_kernels):
424
+
425
+ if xs is None:
426
+ xs = self.resblocks[i*self.num_kernels+j](x, s)
427
+ else:
428
+ xs += self.resblocks[i*self.num_kernels+j](x, s)
429
+ x = xs / self.num_kernels
430
+ # x = x + (1 / self.alphas[i+1]) * (torch.sin(self.alphas[i+1] * x) ** 2) # noisy
431
+ x = self.conv_post(x)
432
+ x = torch.tanh(x)
433
+
434
+ return x
435
+
436
+ class AdainResBlk1d(nn.Module):
437
+
438
+ # also used in ProsodyPredictor()
439
+
440
+ def __init__(self, dim_in, dim_out, style_dim=64, actv=nn.LeakyReLU(0.2),
441
+ upsample='none', dropout_p=0.0):
442
+ super().__init__()
443
+ self.actv = actv
444
+ self.upsample_type = upsample
445
+ self.upsample = UpSample1d(upsample)
446
+ self.learned_sc = dim_in != dim_out
447
+ self._build_weights(dim_in, dim_out, style_dim)
448
+ if upsample == 'none':
449
+ self.pool = nn.Identity()
450
+ else:
451
+ self.pool = weight_norm(nn.ConvTranspose1d(
452
+ dim_in, dim_in, kernel_size=3, stride=2, groups=dim_in, padding=1, output_padding=1))
453
+
454
+ def _build_weights(self, dim_in, dim_out, style_dim):
455
+ self.conv1 = weight_norm(nn.Conv1d(dim_in, dim_out, 3, 1, 1))
456
+ self.conv2 = weight_norm(nn.Conv1d(dim_out, dim_out, 3, 1, 1))
457
+ self.norm1 = AdaIN1d(style_dim, dim_in)
458
+ self.norm2 = AdaIN1d(style_dim, dim_out)
459
+ if self.learned_sc:
460
+ self.conv1x1 = weight_norm(
461
+ nn.Conv1d(dim_in, dim_out, 1, 1, 0, bias=False))
462
+
463
+ def _shortcut(self, x):
464
+ x = self.upsample(x)
465
+ if self.learned_sc:
466
+ x = self.conv1x1(x)
467
+ return x
468
+
469
+ def _residual(self, x, s):
470
+ x = self.norm1(x, s)
471
+ x = self.actv(x)
472
+ x = self.pool(x)
473
+ x = self.conv1(x)
474
+ x = self.norm2(x, s)
475
+ x = self.actv(x)
476
+ x = self.conv2(x)
477
+ return x
478
+
479
+ def forward(self, x, s):
480
+ out = self._residual(x, s)
481
+ out = (out + self._shortcut(x)) / math.sqrt(2)
482
+ return out
483
+
484
+
485
+ class UpSample1d(nn.Module):
486
+ def __init__(self, layer_type):
487
+ super().__init__()
488
+ self.layer_type = layer_type
489
+
490
+ def forward(self, x):
491
+ if self.layer_type == 'none':
492
+ return x
493
+ else:
494
+ return F.interpolate(x, scale_factor=2, mode='nearest-exact')
495
+
496
+
497
+ class Decoder(nn.Module):
498
+ def __init__(self, dim_in=512, F0_channel=512, style_dim=64, dim_out=80,
499
+ resblock_kernel_sizes=[3, 7, 11],
500
+ upsample_rates=[10, 5, 3, 2],
501
+ upsample_initial_channel=512,
502
+ resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
503
+ upsample_kernel_sizes=[20, 10, 6, 4]):
504
+ super().__init__()
505
+
506
+ self.decode = nn.ModuleList()
507
+
508
+ self.encode = AdainResBlk1d(dim_in + 2, 1024, style_dim)
509
+
510
+ self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
511
+ self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
512
+ self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
513
+ self.decode.append(AdainResBlk1d(
514
+ 1024 + 2 + 64, 512, style_dim, upsample=True))
515
+
516
+ self.F0_conv = weight_norm(
517
+ nn.Conv1d(1, 1, kernel_size=3, stride=2, groups=1, padding=1)) # smooth
518
+
519
+ self.N_conv = weight_norm(
520
+ nn.Conv1d(1, 1, kernel_size=3, stride=2, groups=1, padding=1))
521
+
522
+ self.asr_res = nn.Sequential(
523
+ weight_norm(nn.Conv1d(512, 64, kernel_size=1)),
524
+ )
525
+
526
+ self.generator = Generator(style_dim, resblock_kernel_sizes, upsample_rates,
527
+ upsample_initial_channel, resblock_dilation_sizes, upsample_kernel_sizes)
528
+
529
+ def forward(self, asr=None, F0_curve=None, N=None, s=None):
530
+
531
+
532
+ F0 = self.F0_conv(F0_curve)
533
+ N = self.N_conv(N)
534
+
535
+
536
+ x = torch.cat([asr, F0, N], axis=1)
537
+
538
+ x = self.encode(x, s)
539
+
540
+ asr_res = self.asr_res(asr)
541
+
542
+ res = True
543
+ for block in self.decode:
544
+ if res:
545
+
546
+ x = torch.cat([x, asr_res, F0, N], axis=1)
547
+
548
+ x = block(x, s)
549
+ if block.upsample_type != "none":
550
+ res = False
551
+
552
+ x = self.generator(x, s, F0_curve)
553
+ return x
554
+
555
+
556
+ class MelSpec(torch.nn.Module):
557
+
558
+ def __init__(self,
559
+ sample_rate=17402, # https://github.com/fakerybakery/styletts2-cli/blob/main/msinference.py = Default 16000. However 17400 vocalises better also "en_US/vctk_p274"
560
+ n_fft=2048,
561
+ win_length=1200,
562
+ hop_length=300,
563
+ n_mels=80
564
+ ):
565
+ '''avoids dependency on torchaudio'''
566
+ super().__init__()
567
+ self.n_fft = n_fft
568
+ self.win_length = win_length if win_length is not None else n_fft
569
+ self.hop_length = hop_length if hop_length is not None else self.win_length // 2
570
+ # --
571
+ f_min = 0.0
572
+ f_max = float(sample_rate // 2)
573
+ all_freqs = torch.linspace(0, sample_rate // 2, n_fft//2+1)
574
+ m_min = 2595.0 * math.log10(1.0 + (f_min / 700.0))
575
+ m_max = 2595.0 * math.log10(1.0 + (f_max / 700.0))
576
+ m_pts = torch.linspace(m_min, m_max, n_mels + 2)
577
+ f_pts = 700.0 * (10 ** (m_pts / 2595.0) - 1.0)
578
+ f_diff = f_pts[1:] - f_pts[:-1] # (n_mels + 1)
579
+ slopes = f_pts.unsqueeze(0) - all_freqs.unsqueeze(1)
580
+ zero = torch.zeros(1)
581
+ down_slopes = (-1.0 * slopes[:, :-2]) / f_diff[:-1] # (n_freqs, n_mels)
582
+ up_slopes = slopes[:, 2:] / f_diff[1:] # (n_freqs, n_mels)
583
+ fb = torch.max(zero, torch.min(down_slopes, up_slopes))
584
+ # --
585
+ self.register_buffer('fb', fb, persistent=False)
586
+ window = torch.hann_window(self.win_length)
587
+ self.register_buffer('window', window, persistent=False)
588
+
589
+ def forward(self, x):
590
+ spec_f = torch.stft(x,
591
+ self.n_fft,
592
+ self.hop_length,
593
+ self.win_length,
594
+ self.window,
595
+ center=True,
596
+ pad_mode="reflect",
597
+ normalized=False,
598
+ onesided=True,
599
+ return_complex=True) # [bs, 1025, 56]
600
+ mel_specgram = torch.matmul(spec_f.abs().pow(2).transpose(1, 2), self.fb).transpose(1, 2)
601
+ return mel_specgram[:, None, :, :] # [bs, 1, 80, time]
602
+
603
+
604
+ class LearnedDownSample(nn.Module):
605
+ def __init__(self, dim_in):
606
+ super().__init__()
607
+ self.conv = spectral_norm(nn.Conv2d(dim_in, dim_in, kernel_size=(
608
+ 3, 3), stride=(2, 2), groups=dim_in, padding=1))
609
+
610
+ def forward(self, x):
611
+ return self.conv(x)
612
+
613
+
614
+ class ResBlk(nn.Module):
615
+ def __init__(self,
616
+ dim_in, dim_out):
617
+ super().__init__()
618
+ self.actv = nn.LeakyReLU(0.2) # .07 also nice
619
+ self.downsample_res = LearnedDownSample(dim_in)
620
+ self.learned_sc = dim_in != dim_out
621
+ self.conv1 = spectral_norm(nn.Conv2d(dim_in, dim_in, 3, 1, 1))
622
+ self.conv2 = spectral_norm(nn.Conv2d(dim_in, dim_out, 3, 1, 1))
623
+ if self.learned_sc:
624
+ self.conv1x1 = spectral_norm(
625
+ nn.Conv2d(dim_in, dim_out, 1, 1, 0, bias=False))
626
+
627
+ def _shortcut(self, x):
628
+ if self.learned_sc:
629
+ x = self.conv1x1(x)
630
+ if x.shape[3] % 2 != 0: # [bs, 128, Freq, Time]
631
+ x = torch.cat([x, x[:, :, :, -1:]], dim=3)
632
+ return F.interpolate(x, scale_factor=.5, mode='nearest-exact') # F.avg_pool2d(x, 2)
633
+
634
+ def _residual(self, x):
635
+ x = self.actv(x)
636
+ x = self.conv1(x)
637
+ x = self.downsample_res(x)
638
+ x = self.actv(x)
639
+ x = self.conv2(x)
640
+ return x
641
+
642
+ def forward(self, x):
643
+ x = self._shortcut(x) + self._residual(x)
644
+ return x / math.sqrt(2) # unit variance
645
+
646
+
647
+ class StyleEncoder(nn.Module):
648
+
649
+ # for both acoustic & prosodic ref_s/p
650
+
651
+ def __init__(self,
652
+ dim_in=64,
653
+ style_dim=128,
654
+ max_conv_dim=512):
655
+ super().__init__()
656
+ blocks = [spectral_norm(nn.Conv2d(1, dim_in, 3, stride=1, padding=1))]
657
+ for _ in range(4):
658
+ dim_out = min(dim_in * 2,
659
+ max_conv_dim)
660
+ blocks += [ResBlk(dim_in, dim_out)]
661
+ dim_in = dim_out
662
+ blocks += [nn.LeakyReLU(0.24), # w/o this activation - produces no speech
663
+ spectral_norm(nn.Conv2d(dim_out, dim_out, 5, stride=1, padding=0)),
664
+ nn.LeakyReLU(0.2) # 0.3 sounds nice
665
+ ]
666
+ self.shared = nn.Sequential(*blocks)
667
+ self.unshared = nn.Linear(dim_out, style_dim)
668
+
669
+ def forward(self, x):
670
+ x = self.shared(x)
671
+ x = x.mean(3, keepdims=True) # comment this line for time varying style vector
672
+ x = x.transpose(1, 3)
673
+ s = self.unshared(x)
674
+ return s
675
+
676
+
677
+ class LinearNorm(torch.nn.Module):
678
+ def __init__(self, in_dim, out_dim, bias=True):
679
+ super().__init__()
680
+ self.linear_layer = torch.nn.Linear(in_dim, out_dim, bias=bias)
681
+
682
+ def forward(self, x):
683
+ return self.linear_layer(x)
684
+
685
+
686
+ class LayerNorm(nn.Module):
687
+ def __init__(self, channels, eps=1e-5):
688
+ super().__init__()
689
+ self.channels = channels
690
+ self.eps = eps
691
+
692
+ self.gamma = nn.Parameter(torch.ones(channels))
693
+ self.beta = nn.Parameter(torch.zeros(channels))
694
+
695
+ def forward(self, x):
696
+ x = x.transpose(1, -1)
697
+ x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
698
+ return x.transpose(1, -1)
699
+
700
+
701
+ class TextEncoder(nn.Module):
702
+ def __init__(self, channels, kernel_size, depth, n_symbols):
703
+ super().__init__()
704
+ self.embedding = nn.Embedding(n_symbols, channels)
705
+ padding = (kernel_size - 1) // 2
706
+ self.cnn = nn.ModuleList()
707
+ for _ in range(depth):
708
+ self.cnn.append(nn.Sequential(
709
+ weight_norm(nn.Conv1d(channels, channels, kernel_size=kernel_size, padding=padding)),
710
+ LayerNorm(channels),
711
+ nn.LeakyReLU(0.24))
712
+ )
713
+ self.lstm = nn.LSTM(channels, channels//2, 1,
714
+ batch_first=True, bidirectional=True)
715
+
716
+ def forward(self, x):
717
+ x = self.embedding(x) # [B, T, emb]
718
+ x = x.transpose(1, 2)
719
+ for c in self.cnn:
720
+ x = c(x)
721
+ x = x.transpose(1, 2)
722
+ x, _ = self.lstm(x)
723
+ return x
724
+
725
+
726
+ class AdaLayerNorm(nn.Module):
727
+
728
+ def __init__(self, style_dim, channels=None, eps=1e-5):
729
+ super().__init__()
730
+ self.eps = eps
731
+ self.fc = nn.Linear(style_dim, 1024)
732
+
733
+ def forward(self, x, s):
734
+ h = self.fc(s)
735
+ gamma = h[:, :, :512]
736
+ beta = h[:, :, 512:1024]
737
+ x = F.layer_norm(x, (512, ), eps=self.eps)
738
+ x = (1 + gamma) * x + beta
739
+ return x # [1, 75, 512]
740
+
741
+
742
+ class ProsodyPredictor(nn.Module):
743
+
744
+ def __init__(self, style_dim, d_hid, nlayers, max_dur=50):
745
+ super().__init__()
746
+
747
+ self.text_encoder = DurationEncoder(sty_dim=style_dim,
748
+ d_model=d_hid,
749
+ nlayers=nlayers) # called outside forward
750
+ self.lstm = nn.LSTM(d_hid + style_dim, d_hid // 2,
751
+ 1, batch_first=True, bidirectional=True)
752
+ self.duration_proj = LinearNorm(d_hid, max_dur)
753
+ self.shared = nn.LSTM(d_hid + style_dim, d_hid //
754
+ 2, 1, batch_first=True, bidirectional=True)
755
+ self.F0 = nn.ModuleList([
756
+ AdainResBlk1d(d_hid, d_hid, style_dim),
757
+ AdainResBlk1d(d_hid, d_hid // 2, style_dim, upsample=True),
758
+ AdainResBlk1d(d_hid // 2, d_hid // 2, style_dim),
759
+ ])
760
+ self.N = nn.ModuleList([
761
+ AdainResBlk1d(d_hid, d_hid, style_dim),
762
+ AdainResBlk1d(d_hid, d_hid // 2, style_dim, upsample=True),
763
+ AdainResBlk1d(d_hid // 2, d_hid // 2, style_dim)
764
+ ])
765
+ self.F0_proj = nn.Conv1d(d_hid // 2, 1, 1, 1, 0)
766
+ self.N_proj = nn.Conv1d(d_hid // 2, 1, 1, 1, 0)
767
+
768
+ def F0Ntrain(self, x, s):
769
+
770
+ x, _ = self.shared(x) # [bs, time, ch] LSTM
771
+
772
+ x = x.transpose(1, 2) # [bs, ch, time]
773
+
774
+ F0 = x
775
+
776
+ for block in self.F0:
777
+ # print(f'LOOP {F0.shape=} {s.shape=}\n')
778
+ # )N F0.shape=torch.Size([1, 512, 147]) s.shape=torch.Size([1, 128])
779
+ # This is an AdainResBlk1d expects conv1d dimensions
780
+ F0 = block(F0, s)
781
+ F0 = self.F0_proj(F0)
782
+
783
+ N = x
784
+
785
+ for block in self.N:
786
+ N = block(N, s)
787
+ N = self.N_proj(N)
788
+
789
+ return F0, N
790
+
791
+ def forward(self, d_en=None, s=None):
792
+ blend = self.text_encoder(d_en, s)
793
+ x, _ = self.lstm(blend)
794
+ dur = self.duration_proj(x) # [bs, 150, 50]
795
+
796
+ _, input_length, classifier_50 = dur.shape
797
+
798
+ dur = dur[0, :, :]
799
+ dur = torch.sigmoid(dur).sum(1)
800
+ dur = dur.round().clamp(min=1).to(torch.int64)
801
+ aln_trg = torch.zeros(1,
802
+ dur.sum(),
803
+ input_length,
804
+ device=s.device)
805
+ c_frame = 0
806
+ for i in range(input_length):
807
+ aln_trg[:, c_frame:c_frame + dur[i], i] = 1
808
+ c_frame += dur[i]
809
+ en = torch.bmm(aln_trg, blend)
810
+ F0_pred, N_pred = self.F0Ntrain(en, s)
811
+ return aln_trg, F0_pred, N_pred
812
+
813
+
814
+ class DurationEncoder(nn.Module):
815
+
816
+ def __init__(self, sty_dim=128, d_model=512, nlayers=3):
817
+ super().__init__()
818
+ self.lstms = nn.ModuleList()
819
+ for _ in range(nlayers):
820
+ self.lstms.append(nn.LSTM(d_model + sty_dim,
821
+ d_model // 2,
822
+ num_layers=1,
823
+ batch_first=True,
824
+ bidirectional=True
825
+ ))
826
+ self.lstms.append(AdaLayerNorm(sty_dim, d_model))
827
+
828
+
829
+ def forward(self, x, style):
830
+
831
+ _, _, input_lengths = x.shape # [bs, 512, time]
832
+
833
+ style = _tile(style, length=x.shape[2]).transpose(1, 2)
834
+ x = x.transpose(1, 2)
835
+
836
+ for block in self.lstms:
837
+ if isinstance(block, AdaLayerNorm):
838
+
839
+ x = block(x, style) # LSTM has transposed x
840
+
841
+ else:
842
+ x = torch.cat([x, style], axis=2)
843
+ # LSTM
844
+
845
+ x,_ = block(x) # expects [bs, time, chan] OUTPUTS [bs, time, 2*chan] 2x FROM BIDIRECTIONAL
846
+
847
+ return torch.cat([x, style], axis=2) # predictor.lstm()