-
Notifications
You must be signed in to change notification settings - Fork 82
/
gen.py
340 lines (280 loc) · 13.9 KB
/
gen.py
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
# coding: utf-8
import numpy as np
import torch
import pysptk
import pyworld
import librosa
from sklearn.preprocessing import MinMaxScaler
from nnmnkwii.io import hts
from nnmnkwii.frontend import merlin as fe
from nnmnkwii.postfilters import merlin_post_filter
from nnmnkwii.preprocessing.f0 import interp1d
from nnsvs.io.hts import get_note_indices
from nnsvs.multistream import multi_stream_mlpg, get_static_stream_sizes
from nnsvs.multistream import select_streams, split_streams
from nnsvs.base import PredictionType
def get_windows(num_window=1):
windows = [(0, 0, np.array([1.0]))]
if num_window >= 2:
windows.append((1, 1, np.array([-0.5, 0.0, 0.5])))
if num_window >= 3:
windows.append((1, 1, np.array([1.0, -2.0, 1.0])))
if num_window >= 4:
raise ValueError(f"Not supported num windows: {num_window}")
return windows
def _midi_to_hz(x, idx, log_f0=False):
z = np.zeros(len(x))
indices = x[:, idx] > 0
z[indices] = librosa.midi_to_hz(x[indices, idx])
if log_f0:
z[indices] = np.log(z[indices])
return z
def _is_silence(l):
is_full_context = "@" in l
if is_full_context:
is_silence = ("-sil" in l or "-pau" in l)
else:
is_silence = (l == "sil" or l == "pau")
return is_silence
def predict_timelag(device, labels, timelag_model, timelag_config, timelag_in_scaler, timelag_out_scaler,
binary_dict, continuous_dict,
pitch_indices=None, log_f0_conditioning=True,
allowed_range=[-20, 20], allowed_range_rest=[-40, 40]):
# round start/end times just in case.
labels.round_()
# Extract note-level labels
note_indices = get_note_indices(labels)
note_labels = labels[note_indices]
# Extract musical/linguistic context
timelag_linguistic_features = fe.linguistic_features(
note_labels, binary_dict, continuous_dict,
add_frame_features=False, subphone_features=None).astype(np.float32)
# Adjust input features if we use log-f0 conditioning
if log_f0_conditioning:
if pitch_indices is None:
raise ValueError("Pitch feature indices must be specified!")
for idx in pitch_indices:
timelag_linguistic_features[:, idx] = interp1d(
_midi_to_hz(timelag_linguistic_features, idx, log_f0_conditioning),
kind="slinear")
# Normalization
timelag_linguistic_features = timelag_in_scaler.transform(timelag_linguistic_features)
if isinstance(timelag_in_scaler, MinMaxScaler):
# clip to feature range
timelag_linguistic_features = np.clip(
timelag_linguistic_features, timelag_in_scaler.feature_range[0],
timelag_in_scaler.feature_range[1])
# Run model
x = torch.from_numpy(timelag_linguistic_features).unsqueeze(0).to(device)
# Run model
if timelag_model.prediction_type() == PredictionType.PROBABILISTIC:
# (B, T, D_out)
max_mu, max_sigma = timelag_model.inference(x, [x.shape[1]])
if np.any(timelag_config.has_dynamic_features):
# Apply denormalization
# (B, T, D_out) -> (T, D_out)
max_sigma_sq = max_sigma.squeeze(0).cpu().data.numpy() ** 2 * timelag_out_scaler.var_
max_mu = timelag_out_scaler.inverse_transform(max_mu.squeeze(0).cpu().data.numpy())
# (T, D_out) -> (T, static_dim)
pred_timelag = multi_stream_mlpg(max_mu, max_sigma_sq, get_windows(timelag_config.num_windows),
timelag_config.stream_sizes, timelag_config.has_dynamic_features)
else:
# Apply denormalization
pred_timelag = timelag_out_scaler.inverse_transform(max_mu.squeeze(0).cpu().data.numpy())
else:
# (T, D_out)
pred_timelag = timelag_model.inference(x, [x.shape[1]]).squeeze(0).cpu().data.numpy()
# Apply denormalization
pred_timelag = timelag_out_scaler.inverse_transform(pred_timelag)
if np.any(timelag_config.has_dynamic_features):
# (T, D_out) -> (T, static_dim)
pred_timelag = multi_stream_mlpg(
pred_timelag, timelag_out_scaler.var_, get_windows(timelag_config.num_windows),
timelag_config.stream_sizes, timelag_config.has_dynamic_features)
# Rounding
pred_timelag = np.round(pred_timelag)
# Clip to the allowed range
for idx in range(len(pred_timelag)):
if _is_silence(note_labels.contexts[idx]):
pred_timelag[idx] = np.clip(pred_timelag[idx], allowed_range_rest[0], allowed_range_rest[1])
else:
pred_timelag[idx] = np.clip(pred_timelag[idx], allowed_range[0], allowed_range[1])
# frames -> 100 ns
pred_timelag *= 50000
return pred_timelag
def postprocess_duration(labels, pred_durations, lag):
note_indices = get_note_indices(labels)
# append the end of note
note_indices.append(len(labels))
output_labels = hts.HTSLabelFile()
for i in range(1, len(note_indices)):
# Apply time lag
p = labels[note_indices[i-1]:note_indices[i]]
p.start_times = np.minimum(
np.asarray(p.start_times) + lag[i-1].reshape(-1),
np.asarray(p.end_times) - 50000 * len(p))
p.start_times = np.maximum(p.start_times, 0)
if len(output_labels) > 0:
p.start_times = np.maximum(p.start_times, output_labels.start_times[-1] + 50000)
# Compute normalized phoneme durations
d = fe.duration_features(p)
d_hat = pred_durations[note_indices[i-1]:note_indices[i]]
d_norm = d[0] * d_hat / d_hat.sum()
d_norm = np.round(d_norm)
d_norm[d_norm <= 0] = 1
# TODO: better way to adjust?
if d_norm.sum() != d[0]:
d_norm[-1] += d[0] - d_norm.sum()
p.set_durations(d_norm)
if len(output_labels) > 0:
output_labels.end_times[-1] = p.start_times[0]
for n in p:
output_labels.append(n)
return output_labels
def predict_duration(device, labels, duration_model, duration_config, duration_in_scaler, duration_out_scaler,
lag, binary_dict, continuous_dict, pitch_indices=None, log_f0_conditioning=True):
# Extract musical/linguistic features
duration_linguistic_features = fe.linguistic_features(
labels, binary_dict, continuous_dict,
add_frame_features=False, subphone_features=None).astype(np.float32)
if log_f0_conditioning:
for idx in pitch_indices:
duration_linguistic_features[:, idx] = interp1d(
_midi_to_hz(duration_linguistic_features, idx, log_f0_conditioning),
kind="slinear")
# Apply normalization
duration_linguistic_features = duration_in_scaler.transform(duration_linguistic_features)
if isinstance(duration_in_scaler, MinMaxScaler):
# clip to feature range
duration_linguistic_features = np.clip(
duration_linguistic_features, duration_in_scaler.feature_range[0],
duration_in_scaler.feature_range[1])
# Apply model
x = torch.from_numpy(duration_linguistic_features).float().to(device)
x = x.view(1, -1, x.size(-1))
if duration_model.prediction_type() == PredictionType.PROBABILISTIC:
# (B, T, D_out)
max_mu, max_sigma = duration_model.inference(x, [x.shape[1]])
if np.any(duration_config.has_dynamic_features):
# Apply denormalization
# (B, T, D_out) -> (T, D_out)
max_sigma_sq = max_sigma.squeeze(0).cpu().data.numpy() ** 2 * duration_out_scaler.var_
max_mu = duration_out_scaler.inverse_transform(max_mu.squeeze(0).cpu().data.numpy())
# (T, D_out) -> (T, static_dim)
pred_durations = multi_stream_mlpg(max_mu, max_sigma_sq, get_windows(duration_config.num_windows),
duration_config.stream_sizes, duration_config.has_dynamic_features)
else:
# Apply denormalization
pred_durations = duration_out_scaler.inverse_transform(max_mu.squeeze(0).cpu().data.numpy())
else:
# (T, D_out)
pred_durations = duration_model.inference(x, [x.shape[1]]).squeeze(0).cpu().data.numpy()
# Apply denormalization
pred_durations = duration_out_scaler.inverse_transform(pred_durations)
if np.any(duration_config.has_dynamic_features):
# (T, D_out) -> (T, static_dim)
pred_durations = multi_stream_mlpg(
pred_durations, duration_out_scaler.var_, get_windows(duration_config.num_windows),
duration_config.stream_sizes, duration_config.has_dynamic_features)
pred_durations[pred_durations <= 0] = 1
pred_durations = np.round(pred_durations)
return pred_durations
def predict_acoustic(device, labels, acoustic_model, acoustic_config, acoustic_in_scaler,
acoustic_out_scaler, binary_dict, continuous_dict,
subphone_features="coarse_coding",
pitch_indices=None, log_f0_conditioning=True):
# Musical/linguistic features
linguistic_features = fe.linguistic_features(labels,
binary_dict, continuous_dict,
add_frame_features=True,
subphone_features=subphone_features)
if log_f0_conditioning:
for idx in pitch_indices:
linguistic_features[:, idx] = interp1d(
_midi_to_hz(linguistic_features, idx, log_f0_conditioning),
kind="slinear")
# Apply normalization
linguistic_features = acoustic_in_scaler.transform(linguistic_features)
if isinstance(acoustic_in_scaler, MinMaxScaler):
# clip to feature range
linguistic_features = np.clip(
linguistic_features, acoustic_in_scaler.feature_range[0],
acoustic_in_scaler.feature_range[1])
# Predict acoustic features
x = torch.from_numpy(linguistic_features).float().to(device)
x = x.view(1, -1, x.size(-1))
if acoustic_model.prediction_type() == PredictionType.PROBABILISTIC:
# (B, T, D_out)
max_mu, max_sigma = acoustic_model.inference(x, [x.shape[1]])
if np.any(acoustic_config.has_dynamic_features):
# Apply denormalization
# (B, T, D_out) -> (T, D_out)
max_sigma_sq = max_sigma.squeeze(0).cpu().data.numpy() ** 2 * acoustic_out_scaler.var_
max_mu = acoustic_out_scaler.inverse_transform(max_mu.squeeze(0).cpu().data.numpy())
# (T, D_out) -> (T, static_dim)
pred_acoustic = multi_stream_mlpg(max_mu, max_sigma_sq, get_windows(acoustic_config.num_windows),
acoustic_config.stream_sizes, acoustic_config.has_dynamic_features)
else:
# Apply denormalization
pred_acoustic = acoustic_out_scaler.inverse_transform(max_mu.squeeze(0).cpu().data.numpy())
else:
# (T, D_out)
pred_acoustic = acoustic_model.inference(x, [x.shape[1]]).squeeze(0).cpu().data.numpy()
# Apply denormalization
pred_acoustic = acoustic_out_scaler.inverse_transform(pred_acoustic)
if np.any(acoustic_config.has_dynamic_features):
# (T, D_out) -> (T, static_dim)
pred_acoustic = multi_stream_mlpg(
pred_acoustic, acoustic_out_scaler.var_, get_windows(acoustic_config.num_windows),
acoustic_config.stream_sizes, acoustic_config.has_dynamic_features)
return pred_acoustic
def gen_waveform(labels, acoustic_features,
binary_dict, continuous_dict, stream_sizes, has_dynamic_features,
subphone_features="coarse_coding", log_f0_conditioning=True, pitch_idx=None,
num_windows=3, post_filter=True, sample_rate=48000, frame_period=5,
relative_f0=True):
windows = get_windows(num_windows)
# Apply MLPG if necessary
if np.any(has_dynamic_features):
static_stream_sizes = get_static_stream_sizes(
stream_sizes, has_dynamic_features, len(windows))
else:
static_stream_sizes = stream_sizes
# Split multi-stream features
mgc, target_f0, vuv, bap = split_streams(acoustic_features, static_stream_sizes)
# Gen waveform by the WORLD vocodoer
fftlen = pyworld.get_cheaptrick_fft_size(sample_rate)
alpha = pysptk.util.mcepalpha(sample_rate)
if post_filter:
mgc = merlin_post_filter(mgc, alpha)
spectrogram = pysptk.mc2sp(mgc, fftlen=fftlen, alpha=alpha)
aperiodicity = pyworld.decode_aperiodicity(bap.astype(np.float64), sample_rate, fftlen)
# fill aperiodicity with ones for unvoiced regions
aperiodicity[vuv.reshape(-1) < 0.5, :] = 1.0
# WORLD fails catastrophically for out of range aperiodicity
aperiodicity = np.clip(aperiodicity, 0.0, 1.0)
### F0 ###
if relative_f0:
diff_lf0 = target_f0
# need to extract pitch sequence from the musical score
linguistic_features = fe.linguistic_features(labels,
binary_dict, continuous_dict,
add_frame_features=True,
subphone_features=subphone_features)
f0_score = _midi_to_hz(linguistic_features, pitch_idx, False)[:, None]
lf0_score = f0_score.copy()
nonzero_indices = np.nonzero(lf0_score)
lf0_score[nonzero_indices] = np.log(f0_score[nonzero_indices])
lf0_score = interp1d(lf0_score, kind="slinear")
f0 = diff_lf0 + lf0_score
f0[vuv < 0.5] = 0
f0[np.nonzero(f0)] = np.exp(f0[np.nonzero(f0)])
else:
f0 = target_f0
f0[vuv < 0.5] = 0
f0[np.nonzero(f0)] = np.exp(f0[np.nonzero(f0)])
generated_waveform = pyworld.synthesize(f0.flatten().astype(np.float64),
spectrogram.astype(np.float64),
aperiodicity.astype(np.float64),
sample_rate, frame_period)
return generated_waveform