"""
Use a triplet-loss to learn a similarity metric between short spectrograms
UNSUPERVISED LEARNING OF SEMANTIC AUDIO REPRESENTATIONS
https://arxiv.org/pdf/1711.02209.pdf
"""
import numpy as np
import zounds
from zounds.spectral import apply_scale
samplerate = zounds.SR11025()
BaseModel = zounds.resampled(resample_to=samplerate, store_resampled=True)
scale_bands = 96
spectrogram_duration = 64
anchor_slice = slice(spectrogram_duration, spectrogram_duration * 2)
scale = zounds.GeometricScale(
start_center_hz=50,
stop_center_hz=samplerate.nyquist,
bandwidth_ratio=0.115,
n_bands=scale_bands)
scale.ensure_overlap_ratio()
spectrogram_duration = 64
windowing_scheme = zounds.HalfLapped()
spectrogram_sample_rate = zounds.SampleRate(
frequency=windowing_scheme.frequency * (spectrogram_duration // 2),
import zounds
Resampled = zounds.resampled(resample_to=zounds.SR11025())
@zounds.simple_in_memory_settings
class Sound(Resampled):
"""
A simple pipeline that computes a perceptually weighted modified discrete
cosine transform, and "persists" feature data in an in-memory store.
"""
windowed = zounds.ArrayWithUnitsFeature(
zounds.SlidingWindow,
needs=Resampled.resampled,
wscheme=zounds.HalfLapped(),
wfunc=zounds.OggVorbisWindowingFunc(),
store=True)
mdct = zounds.ArrayWithUnitsFeature(
zounds.MDCT,
needs=windowed)
weighted = zounds.ArrayWithUnitsFeature(
lambda x: x * zounds.AWeighting(),
needs=mdct)
if __name__ == '__main__':
# produce some audio to test our pipeline, and encode it as FLAC
synth = zounds.SineSynthesizer(zounds.SR44100())
import zounds
Resampled = zounds.resampled(resample_to=zounds.SR11025())
@zounds.simple_in_memory_settings
class Sound(Resampled):
"""
A simple pipeline that computes a perceptually weighted modified discrete
cosine transform, and "persists" feature data in an in-memory store.
"""
windowed = zounds.ArrayWithUnitsFeature(
zounds.SlidingWindow,
needs=Resampled.resampled,
wscheme=zounds.HalfLapped(),
wfunc=zounds.OggVorbisWindowingFunc(),
store=True)
mdct = zounds.ArrayWithUnitsFeature(zounds.MDCT, needs=windowed)
weighted = zounds.ArrayWithUnitsFeature(lambda x: x * zounds.AWeighting(),
needs=mdct)
if __name__ == '__main__':
# produce some audio to test our pipeline, and encode it as FLAC
synth = zounds.SineSynthesizer(zounds.SR44100())
samples = synth.synthesize(zounds.Seconds(5), [220., 440., 880.])
encoded = samples.encode(fmt='FLAC')
import argparse
from random import choice
import featureflow as ff
import numpy as np
from torch import nn
from torch.nn import functional as F
from torch.optim import Adam
import zounds
from zounds.learn import Conv1d, ConvTranspose1d, to_var, from_var
from zounds.timeseries import categorical, inverse_categorical
samplerate = zounds.SR11025()
BaseModel = zounds.resampled(resample_to=samplerate, store_resampled=True)
window_size = 8192
wscheme = zounds.SampleRate(frequency=samplerate.frequency *
(window_size // 2),
duration=samplerate.frequency * window_size)
@zounds.simple_lmdb_settings('ae', map_size=1e10, user_supplied_id=True)
class Sound(BaseModel):
windowed = zounds.ArrayWithUnitsFeature(zounds.SlidingWindow,
wscheme=wscheme,
needs=BaseModel.resampled)
mu_law = zounds.ArrayWithUnitsFeature(zounds.mu_law, needs=windowed)
categorical = zounds.ArrayWithUnitsFeature(categorical, needs=windowed)
import numpy as np
import zounds
from zounds.spectral import apply_scale
samplerate = zounds.SR11025()
BaseModel = zounds.resampled(resample_to=samplerate)
scale_bands = 96
spectrogram_duration = 64
anchor_slice = slice(spectrogram_duration, spectrogram_duration * 2)
scale = zounds.GeometricScale(
start_center_hz=50,
stop_center_hz=samplerate.nyquist,
bandwidth_ratio=0.115,
n_bands=scale_bands)
scale.ensure_overlap_ratio()
spectrogram_duration = 64
windowing_scheme = zounds.HalfLapped()
spectrogram_sample_rate = zounds.SampleRate(
frequency=windowing_scheme.frequency * (spectrogram_duration // 2),
duration=windowing_scheme.frequency * spectrogram_duration)
def spectrogram(x):
x = apply_scale(
np.abs(x.real), scale, window=zounds.OggVorbisWindowingFunc())
x = zounds.log_modulus(x * 100)