forked from magenta/ddsp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
synths.py
280 lines (232 loc) · 9.66 KB
/
synths.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
# Copyright 2020 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
"""Library of synthesizer functions."""
from ddsp import core
from ddsp import processors
import gin
import tensorflow.compat.v2 as tf
# TODO(jesseengel): Rename Additive to Harmonic.
@gin.register
class Additive(processors.Processor):
"""Synthesize audio with a bank of harmonic sinusoidal oscillators."""
def __init__(self,
n_samples=64000,
sample_rate=16000,
scale_fn=core.exp_sigmoid,
normalize_below_nyquist=True,
name='additive'):
super().__init__(name=name)
self.n_samples = n_samples
self.sample_rate = sample_rate
self.scale_fn = scale_fn
self.normalize_below_nyquist = normalize_below_nyquist
def get_controls(self,
amplitudes,
harmonic_distribution,
f0_hz):
"""Convert network output tensors into a dictionary of synthesizer controls.
Args:
amplitudes: 3-D Tensor of synthesizer controls, of shape
[batch, time, 1].
harmonic_distribution: 3-D Tensor of synthesizer controls, of shape
[batch, time, n_harmonics].
f0_hz: Fundamental frequencies in hertz. Shape [batch, time, 1].
Returns:
controls: Dictionary of tensors of synthesizer controls.
"""
# Scale the amplitudes.
if self.scale_fn is not None:
amplitudes = self.scale_fn(amplitudes)
harmonic_distribution = self.scale_fn(harmonic_distribution)
# Bandlimit the harmonic distribution.
if self.normalize_below_nyquist:
n_harmonics = int(harmonic_distribution.shape[-1])
harmonic_frequencies = core.get_harmonic_frequencies(f0_hz,
n_harmonics)
harmonic_distribution = core.remove_above_nyquist(harmonic_frequencies,
harmonic_distribution,
self.sample_rate)
# Normalize
harmonic_distribution /= tf.reduce_sum(harmonic_distribution,
axis=-1,
keepdims=True)
return {'amplitudes': amplitudes,
'harmonic_distribution': harmonic_distribution,
'f0_hz': f0_hz}
def get_signal(self, amplitudes, harmonic_distribution, f0_hz):
"""Synthesize audio with additive synthesizer from controls.
Args:
amplitudes: Amplitude tensor of shape [batch, n_frames, 1]. Expects
float32 that is strictly positive.
harmonic_distribution: Tensor of shape [batch, n_frames, n_harmonics].
Expects float32 that is strictly positive and normalized in the last
dimension.
f0_hz: The fundamental frequency in Hertz. Tensor of shape [batch,
n_frames, 1].
Returns:
signal: A tensor of harmonic waves of shape [batch, n_samples].
"""
signal = core.harmonic_synthesis(
frequencies=f0_hz,
amplitudes=amplitudes,
harmonic_distribution=harmonic_distribution,
n_samples=self.n_samples,
sample_rate=self.sample_rate)
return signal
@gin.register
class FilteredNoise(processors.Processor):
"""Synthesize audio by filtering white noise."""
def __init__(self,
n_samples=64000,
window_size=257,
scale_fn=core.exp_sigmoid,
initial_bias=-5.0,
name='filtered_noise'):
super().__init__(name=name)
self.n_samples = n_samples
self.window_size = window_size
self.scale_fn = scale_fn
self.initial_bias = initial_bias
def get_controls(self, magnitudes):
"""Convert network outputs into a dictionary of synthesizer controls.
Args:
magnitudes: 3-D Tensor of synthesizer parameters, of shape [batch, time,
n_filter_banks].
Returns:
controls: Dictionary of tensors of synthesizer controls.
"""
# Scale the magnitudes.
if self.scale_fn is not None:
magnitudes = self.scale_fn(magnitudes + self.initial_bias)
return {'magnitudes': magnitudes}
def get_signal(self, magnitudes):
"""Synthesize audio with filtered white noise.
Args:
magnitudes: Magnitudes tensor of shape [batch, n_frames, n_filter_banks].
Expects float32 that is strictly positive.
Returns:
signal: A tensor of harmonic waves of shape [batch, n_samples, 1].
"""
batch_size = int(magnitudes.shape[0])
signal = tf.random.uniform(
[batch_size, self.n_samples], minval=-1.0, maxval=1.0)
return core.frequency_filter(signal,
magnitudes,
window_size=self.window_size)
@gin.register
class Wavetable(processors.Processor):
"""Synthesize audio from a series of wavetables."""
def __init__(self,
n_samples=64000,
sample_rate=16000,
scale_fn=core.exp_sigmoid,
name='wavetable'):
super().__init__(name=name)
self.n_samples = n_samples
self.sample_rate = sample_rate
self.scale_fn = scale_fn
def get_controls(self,
amplitudes,
wavetables,
f0_hz):
"""Convert network output tensors into a dictionary of synthesizer controls.
Args:
amplitudes: 3-D Tensor of synthesizer controls, of shape
[batch, time, 1].
wavetables: 3-D Tensor of synthesizer controls, of shape
[batch, time, n_harmonics].
f0_hz: Fundamental frequencies in hertz. Shape [batch, time, 1].
Returns:
controls: Dictionary of tensors of synthesizer controls.
"""
# Scale the amplitudes.
if self.scale_fn is not None:
amplitudes = self.scale_fn(amplitudes)
wavetables = self.scale_fn(wavetables)
return {'amplitudes': amplitudes,
'wavetables': wavetables,
'f0_hz': f0_hz}
def get_signal(self, amplitudes, wavetables, f0_hz):
"""Synthesize audio with additive synthesizer from controls.
Args:
amplitudes: Amplitude tensor of shape [batch, n_frames, 1]. Expects
float32 that is strictly positive.
wavetables: Tensor of shape [batch, n_frames, n_wavetable].
f0_hz: The fundamental frequency in Hertz. Tensor of shape [batch,
n_frames, 1].
Returns:
signal: A tensor of of shape [batch, n_samples].
"""
wavetables = core.resample(wavetables, self.n_samples)
signal = core.wavetable_synthesis(amplitudes=amplitudes,
wavetables=wavetables,
frequencies=f0_hz,
n_samples=self.n_samples,
sample_rate=self.sample_rate)
return signal
@gin.register
class Sinusoidal(processors.Processor):
"""Synthesize audio with a bank of arbitrary sinusoidal oscillators."""
def __init__(self,
n_samples=64000,
sample_rate=16000,
amp_scale_fn=core.exp_sigmoid,
amp_resample_method='window',
freq_scale_fn=core.frequencies_sigmoid,
name='sinusoidal'):
super().__init__(name=name)
self.n_samples = n_samples
self.sample_rate = sample_rate
self.amp_scale_fn = amp_scale_fn
self.amp_resample_method = amp_resample_method
self.freq_scale_fn = freq_scale_fn
def get_controls(self, amplitudes, frequencies):
"""Convert network output tensors into a dictionary of synthesizer controls.
Args:
amplitudes: 3-D Tensor of synthesizer controls, of shape
[batch, time, n_sinusoids].
frequencies: 3-D Tensor of synthesizer controls, of shape
[batch, time, n_sinusoids]. Expects strictly positive in Hertz.
Returns:
controls: Dictionary of tensors of synthesizer controls.
"""
# Scale the inputs.
if self.amp_scale_fn is not None:
amplitudes = self.amp_scale_fn(amplitudes)
if self.freq_scale_fn is not None:
frequencies = self.freq_scale_fn(frequencies)
amplitudes = core.remove_above_nyquist(frequencies,
amplitudes,
self.sample_rate)
return {'amplitudes': amplitudes,
'frequencies': frequencies}
def get_signal(self, amplitudes, frequencies):
"""Synthesize audio with sinusoidal synthesizer from controls.
Args:
amplitudes: Amplitude tensor of shape [batch, n_frames, n_sinusoids].
Expects float32 that is strictly positive.
frequencies: Tensor of shape [batch, n_frames, n_sinusoids].
Expects float32 in Hertz that is strictly positive.
Returns:
signal: A tensor of harmonic waves of shape [batch, n_samples].
"""
# Create sample-wise envelopes.
amplitude_envelopes = core.resample(amplitudes, self.n_samples,
method=self.amp_resample_method)
frequency_envelopes = core.resample(frequencies, self.n_samples)
signal = core.oscillator_bank(frequency_envelopes=frequency_envelopes,
amplitude_envelopes=amplitude_envelopes,
sample_rate=self.sample_rate)
return signal