-
Notifications
You must be signed in to change notification settings - Fork 16
/
speech_tools.py
240 lines (167 loc) · 7.29 KB
/
speech_tools.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
import librosa
import numpy as np
import pyworld
import glob
import pickle
from tqdm import tqdm
def load_wavs(wav_dir, sr):
wavs = list()
for file in glob.glob(wav_dir + '/*.wav'):
wav, _ = librosa.load(file, sr=sr, mono=True)
wavs.append(wav)
return wavs
def world_decompose(wav, fs, frame_period=5.0):
# Decompose speech signal into f0, spectral envelope and aperiodicity using WORLD
wav = wav.astype(np.float64)
f0, timeaxis = pyworld.harvest(wav, fs, frame_period=frame_period, f0_floor=71.0, f0_ceil=800.0)
sp = pyworld.cheaptrick(wav, f0, timeaxis, fs)
ap = pyworld.d4c(wav, f0, timeaxis, fs)
return f0, timeaxis, sp, ap
def world_encode_spectral_envelop(sp, fs, dim=24):
# Get Mel-cepstral coefficients (MCEPs)
# sp = sp.astype(np.float64)
coded_sp = pyworld.code_spectral_envelope(sp, fs, dim)
return coded_sp
def world_decode_spectral_envelop(coded_sp, fs):
fftlen = pyworld.get_cheaptrick_fft_size(fs)
# coded_sp = coded_sp.astype(np.float32)
# coded_sp = np.ascontiguousarray(coded_sp)
decoded_sp = pyworld.decode_spectral_envelope(coded_sp, fs, fftlen)
return decoded_sp
def world_encode_data(wavs, fs, frame_period=5.0, coded_dim=24):
f0s = list()
timeaxes = list()
sps = list()
aps = list()
coded_sps = list()
for wav in tqdm(wavs):
f0, timeaxis, sp, ap = world_decompose(wav=wav, fs=fs, frame_period=frame_period)
coded_sp = world_encode_spectral_envelop(sp=sp, fs=fs, dim=coded_dim)
f0s.append(f0)
timeaxes.append(timeaxis)
sps.append(sp)
aps.append(ap)
coded_sps.append(coded_sp)
return f0s, timeaxes, sps, aps, coded_sps
def transpose_in_list(lst):
transposed_lst = list()
for array in lst:
transposed_lst.append(array.T)
return transposed_lst
def world_decode_data(coded_sps, fs):
decoded_sps = list()
for coded_sp in coded_sps:
decoded_sp = world_decode_spectral_envelop(coded_sp, fs)
decoded_sps.append(decoded_sp)
return decoded_sps
def world_speech_synthesis(f0, decoded_sp, ap, fs, frame_period):
# decoded_sp = decoded_sp.astype(np.float64)
wav = pyworld.synthesize(f0, decoded_sp, ap, fs, frame_period)
# Librosa could not save wav if not doing so
wav = wav.astype(np.float32)
return wav
def world_synthesis_data(f0s, decoded_sps, aps, fs, frame_period):
wavs = list()
for f0, decoded_sp, ap in zip(f0s, decoded_sps, aps):
wav = world_speech_synthesis(f0, decoded_sp, ap, fs, frame_period)
wavs.append(wav)
return wavs
def coded_sps_normalization_fit_transoform(coded_sps):
coded_sps_concatenated = np.concatenate(coded_sps, axis=1)
coded_sps_mean = np.mean(coded_sps_concatenated, axis=1, keepdims=True)
coded_sps_std = np.std(coded_sps_concatenated, axis=1, keepdims=True)
coded_sps_normalized = list()
for coded_sp in coded_sps:
coded_sps_normalized.append((coded_sp - coded_sps_mean) / coded_sps_std)
return coded_sps_normalized, coded_sps_mean, coded_sps_std
def coded_sps_normalization_transoform(coded_sps, coded_sps_mean, coded_sps_std):
coded_sps_normalized = list()
for coded_sp in coded_sps:
coded_sps_normalized.append((coded_sp - coded_sps_mean) / coded_sps_std)
return coded_sps_normalized
def coded_sps_normalization_inverse_transoform(normalized_coded_sps, coded_sps_mean, coded_sps_std):
coded_sps = list()
for normalized_coded_sp in normalized_coded_sps:
coded_sps.append(normalized_coded_sp * coded_sps_std + coded_sps_mean)
return coded_sps
def coded_sp_padding(coded_sp, multiple=4):
num_features = coded_sp.shape[0]
num_frames = coded_sp.shape[1]
num_frames_padded = int(np.ceil(num_frames / multiple)) * multiple
num_frames_diff = num_frames_padded - num_frames
num_pad_left = num_frames_diff // 2
num_pad_right = num_frames_diff - num_pad_left
coded_sp_padded = np.pad(coded_sp, ((0, 0), (num_pad_left, num_pad_right)), 'constant', constant_values=0)
return coded_sp_padded
def wav_padding(wav, sr, frame_period, multiple=4):
assert wav.ndim == 1
num_frames = len(wav)
num_frames_padded = int(
(np.ceil((np.floor(num_frames / (sr * frame_period / 1000)) + 1) / multiple + 1) * multiple - 1) * (
sr * frame_period / 1000))
num_frames_diff = num_frames_padded - num_frames
num_pad_left = num_frames_diff // 2
num_pad_right = num_frames_diff - num_pad_left
wav_padded = np.pad(wav, (num_pad_left, num_pad_right), 'constant', constant_values=0)
return wav_padded
def logf0_statistics(f0s):
log_f0s_concatenated = np.ma.log(np.concatenate(f0s))
log_f0s_mean = log_f0s_concatenated.mean()
log_f0s_std = log_f0s_concatenated.std()
return log_f0s_mean, log_f0s_std
def pitch_conversion(f0, mean_log_src, std_log_src, mean_log_target, std_log_target):
# Logarithm Gaussian normalization for Pitch Conversions
f0_converted = np.exp((np.log(f0) - mean_log_src) / std_log_src * std_log_target + mean_log_target)
return f0_converted
def wavs_to_specs(wavs, n_fft=1024, hop_length=None):
stfts = list()
for wav in wavs:
stft = librosa.stft(wav, n_fft=n_fft, hop_length=hop_length)
stfts.append(stft)
return stfts
def wavs_to_mfccs(wavs, sr, n_fft=1024, hop_length=None, n_mels=128, n_mfcc=24):
mfccs = list()
for wav in wavs:
mfcc = librosa.feature.mfcc(y=wav, sr=sr, n_fft=n_fft, hop_length=hop_length, n_mels=n_mels, n_mfcc=n_mfcc)
mfccs.append(mfcc)
return mfccs
def mfccs_normalization(mfccs):
mfccs_concatenated = np.concatenate(mfccs, axis=1)
mfccs_mean = np.mean(mfccs_concatenated, axis=1, keepdims=True)
mfccs_std = np.std(mfccs_concatenated, axis=1, keepdims=True)
mfccs_normalized = list()
for mfcc in mfccs:
mfccs_normalized.append((mfcc - mfccs_mean) / mfccs_std)
return mfccs_normalized, mfccs_mean, mfccs_std
def save_pickle(path, obj):
with open(path, 'wb') as f:
pickle.dump(obj, f)
def load_pickle(path):
with open(path, 'rb') as f:
return pickle.load(f)
def sample_train_data(dataset_A, dataset_B, n_frames=128):
num_samples = min(len(dataset_A), len(dataset_B))
train_data_A_idx = np.arange(len(dataset_A))
train_data_B_idx = np.arange(len(dataset_B))
np.random.shuffle(train_data_A_idx)
np.random.shuffle(train_data_B_idx)
train_data_A_idx_subset = train_data_A_idx[:num_samples]
train_data_B_idx_subset = train_data_B_idx[:num_samples]
train_data_A = list()
train_data_B = list()
for idx_A, idx_B in zip(train_data_A_idx_subset, train_data_B_idx_subset):
data_A = dataset_A[idx_A]
frames_A_total = data_A.shape[1]
assert frames_A_total >= n_frames
start_A = np.random.randint(frames_A_total - n_frames + 1)
end_A = start_A + n_frames
train_data_A.append(data_A[:, start_A:end_A])
data_B = dataset_B[idx_B]
frames_B_total = data_B.shape[1]
assert frames_B_total >= n_frames
start_B = np.random.randint(frames_B_total - n_frames + 1)
end_B = start_B + n_frames
train_data_B.append(data_B[:, start_B:end_B])
train_data_A = np.array(train_data_A)
train_data_B = np.array(train_data_B)
return train_data_A, train_data_B