-
Notifications
You must be signed in to change notification settings - Fork 0
/
synthetic_corpus.py
251 lines (204 loc) · 8.44 KB
/
synthetic_corpus.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
'''The code is adapted from https://github.com/magenta/ddsp/blob/main/ddsp/training/data_preparation/synthetic_data.py
'''
# Copyright 2021 The DDSP Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Lint as: python3
"""Functions to generate self-supervised signal, EXPERIMENTAL."""
import random
import warnings
import ddsp
import numpy as np
import tensorflow.compat.v2 as tf
from scipy.interpolate import interp1d
def flip(p=0.5):
return np.random.uniform() <= p
def uniform_int(minval=0, maxval=10):
return np.random.random_integers(int(minval), int(maxval))
def uniform_float(minval=0.0, maxval=10.0):
return np.random.uniform(float(minval), float(maxval))
def random_blend(length, env_start=1.0, env_end=0.0, exp_max=2.0):
"""Returns a linear mix between two values, with a random curve steepness."""
exp = uniform_float(-exp_max, exp_max)
v = np.linspace(1.0, 0.0, length) ** (2.0 ** exp)
return env_start * v + env_end * (1.0 - v)
def random_harm_dist(n_harmonics=100, low_pass=True, rand_phase=0.0):
"""Create harmonic distribution out of sinusoidal components."""
n_components = uniform_int(1, 20)
smoothness = uniform_float(1.0, 10.0)
coeffs = np.random.rand(n_components)
freqs = np.random.rand(n_components) * n_harmonics / smoothness
v = []
for i in range(n_components):
v_i = (coeffs[i] * np.cos(
np.linspace(0.0, 2.0 * np.pi * freqs[i], n_harmonics) +
uniform_float(0.0, np.pi * 2.0 * rand_phase)))
v.append(v_i)
if low_pass:
v = [v_i * np.linspace(1.0, uniform_float(0.0, 0.5), n_harmonics) **
uniform_float(0.5, 2.0) for v_i in v]
harm_dist = np.sum(np.stack(v), axis=0)
return harm_dist
def running_mean(x, N):
return np.convolve(x, np.ones((N,)) / N)[(N - 1):]
def random_walk(length, smooth_win):
y = np.cumsum(np.random.normal(loc=0., scale=1., size=(length,)))
y = running_mean(y, smooth_win)
return y
def random_walk_blend(length, env_start=1.0, env_end=0.0, exp_max=2.0):
"""Returns a linear mix between two values, with a random curve steepness."""
def _rw():
v = random_walk(length * 10, min(length, 10))
max_idx = np.argmax(v)
min_idx = np.argmin(v)
if max_idx > min_idx:
max_to_min_v = v[min_idx: max_idx][::-1]
N = max_idx - min_idx
else:
max_to_min_v = v[max_idx: min_idx]
N = min_idx - max_idx
min_v = v[min_idx]
max_v = v[max_idx]
return max_to_min_v, N, max_v, min_v
max_to_min_v, N, max_v, min_v = _rw()
while N < 2:
max_to_min_v, N, max_v, min_v = _rw()
interp_func = interp1d(np.arange(N), max_to_min_v)
ids = np.linspace(0, N - 1, length)
max_to_min_v = interp_func(ids)
blend_coef = (max_to_min_v - min_v) / (max_v - min_v)
return env_start * blend_coef + env_end * (1 - blend_coef)
def generate_control(n_batch=1,
n_timesteps=125,
n_harmonics=100,
n_mags=65,
min_note_length=5,
max_note_length=25,
p_silent=0.1,
p_vibrato=0.8,
get_controls=True,
sample_rate=16000):
harm_amp = np.zeros([n_batch, n_timesteps, 1])
harm_dist = np.zeros([n_batch, n_timesteps, n_harmonics])
f0_midi = np.zeros([n_batch, n_timesteps, 1])
mags = np.zeros([n_batch, n_timesteps, n_mags])
for b in range(n_batch):
t_start = 0
while t_start < n_timesteps:
note_length = uniform_int(min_note_length, max_note_length)
t_end = min(t_start + note_length, n_timesteps)
note_length = t_end - t_start
# Silent?
silent = flip(p_silent)
if silent:
# Amplitudes.
ha_slice = harm_amp[b, t_start:t_end, :]
ha_slice -= 10.0
else:
# Amplitudes.
amp_start = uniform_float(-1.0, 3.0)
amp_end = uniform_float(-1.0, 3.0)
if flip(0.5):
amp_blend = random_walk_blend(note_length, amp_start, amp_end)
else:
amp_blend = random_blend(note_length, amp_start, amp_end)
ha_slice = harm_amp[b, t_start:t_end, :]
ha_slice += amp_blend[:, np.newaxis]
# Add some noise.
ha_slice += uniform_float(0.0, 0.1) * np.random.randn(*ha_slice.shape)
# Harmonic Distribution.
low_pass = flip(0.8)
rand_phase = uniform_float(0.0, 0.4)
harm_dist_start = random_harm_dist(n_harmonics,
low_pass=low_pass,
rand_phase=rand_phase)[np.newaxis, :]
harm_dist_end = random_harm_dist(n_harmonics,
low_pass=low_pass,
rand_phase=rand_phase)[np.newaxis, :]
if flip(0.5):
blend = random_walk_blend(note_length, 1.0, 0.0)[:, np.newaxis]
else:
blend = random_blend(note_length, 1.0, 0.0)[:, np.newaxis]
harm_dist_blend = (harm_dist_start * blend +
harm_dist_end * (1.0 - blend))
hd_slice = harm_dist[b, t_start:t_end, :]
hd_slice += harm_dist_blend
# Add some noise.
hd_slice += uniform_float(0.0, 0.5) * np.random.randn(*hd_slice.shape)
if flip(0.5):
mask_idx = np.random.randint(7, 100)
hd_slice[:, mask_idx:] = -np.inf
# Fundamental Frequency.
f0 = uniform_float(24.0, 84.0)
if flip(p_vibrato):
vib_start = uniform_float(0.0, 1.0)
vib_end = uniform_float(0.0, 1.0)
vib_periods = uniform_float(0.0, note_length * 2.0 / min_note_length)
if flip(0.5):
vib_blend = random_walk_blend(note_length, vib_start, vib_end)
else:
vib_blend = random_blend(note_length, vib_start, vib_end)
if flip(0.5):
vib = vib_blend * np.sin(
np.linspace(0.0, 2.0 * np.pi * vib_periods, note_length))
f0_note = f0 + vib
else:
f0_note = f0 + vib_blend
else:
f0_note = f0 * np.ones([note_length])
f0_slice = f0_midi[b, t_start:t_end, :]
f0_slice += f0_note[:, np.newaxis]
# Add some noise.
f0_slice += uniform_float(0.0, 0.1) * np.random.randn(*f0_slice.shape)
# Filtered Noise.
low_pass = flip(0.8)
rand_phase = uniform_float(0.0, 0.4)
mags_start = random_harm_dist(n_mags,
low_pass=low_pass,
rand_phase=rand_phase)[np.newaxis, :]
mags_end = random_harm_dist(n_mags,
low_pass=low_pass,
rand_phase=rand_phase)[np.newaxis, :]
if flip(0.5):
blend = random_walk_blend(note_length, 1.0, 0.0)[:, np.newaxis]
else:
blend = random_blend(note_length, 1.0, 0.0)[:, np.newaxis]
mags_blend = mags_start * blend + mags_end * (1.0 - blend)
mags_slice = mags[b, t_start:t_end, :]
mags_slice += mags_blend
# Add some noise.
mags_slice += uniform_float(0.0, 0.2) * np.random.randn(*mags_slice.shape)
# # Scale.
mags_slice -= uniform_float(1.0, 10.0)
t_start = t_end
if get_controls:
harm_amp = ddsp.core.exp_sigmoid(harm_amp)
harm_amp /= uniform_float(1.0, [2.0, uniform_float(2.0, 10.0)][flip(0.2)])
# Frequencies.
f0_hz = ddsp.core.midi_to_hz(f0_midi)
if get_controls:
harm_dist = tf.nn.softmax(harm_dist)
harm_dist = ddsp.core.remove_above_nyquist(f0_hz, harm_dist, sample_rate=sample_rate)
harm_dist = ddsp.core.safe_divide(
harm_dist, tf.reduce_sum(harm_dist, axis=-1, keepdims=True))
if get_controls:
mags = ddsp.core.exp_sigmoid(mags)
sin_amps, sin_freqs = ddsp.core.harmonic_to_sinusoidal(
harm_amp, harm_dist, f0_hz, sample_rate=sample_rate)
controls = {'harm_amp': harm_amp,
'harm_dist': harm_dist,
'f0_hz': f0_hz,
'sin_amps': sin_amps,
'sin_freqs': sin_freqs,
'noise_magnitudes': mags}
return controls