def mpc_features(frames, n_mpc=32, n_segments=10):
AF = audiofeature.AudioFeatures(22050,
1024,
n_bands=n_mpc,
scale_to_int_range=False,
critical_band_fn=mel.MelMasters.warp_dense,
power_sqr_abs=False,
noise_level=1.0e-4) #1/10000
mpc = AF.loudness(frames)
mpc_segments = audiofeature.segment(mpc, n_segments)
return AF, mpc_segments, n_mpc
def mpc_w_softplus(frames, n_mpc=32, n_segments=10):
AF = audiofeature.AudioFeatures(22050,
1024,
n_bands=n_mpc,
scale_to_int_range=False,
critical_band_fn=mel.MelMasters.warp_dense,
power_sqr_abs=False,
noise_level=1.0e-4) #1/10000
mpc = theano.tensor.log10(
0.01 * theano.tensor.nnet.softplus(100 * AF.audspec(frames)) + 1.0e-4)
mpc_segments = audiofeature.segment(mpc, n_segments)
return AF, mpc_segments, n_mpc
def mpc_w_softplus(frames, n_mpc=32, n_segments=10):
AF = audiofeature.AudioFeatures(
22050,
1024,
n_bands=n_mpc,
scale_to_int_range=False,
critical_band_fn=mel.MelMasters.warp_dense,
power_sqr_abs=False,
noise_level=1.0e-4,
) # 1/10000
mpc = theano.tensor.log10(0.01 * theano.tensor.nnet.softplus(100 * AF.audspec(frames)) + 1.0e-4)
mpc_segments = audiofeature.segment(mpc, n_segments)
return AF, mpc_segments, n_mpc
def mpc_features(frames, n_mpc=32, n_segments=10):
AF = audiofeature.AudioFeatures(
22050,
1024,
n_bands=n_mpc,
scale_to_int_range=False,
critical_band_fn=mel.MelMasters.warp_dense,
power_sqr_abs=False,
noise_level=1.0e-4,
) # 1/10000
mpc = AF.loudness(frames)
mpc_segments = audiofeature.segment(mpc, n_segments)
return AF, mpc_segments, n_mpc
def mfcc_features(frames, n_mfcc=16, n_segments=10):
af = audiofeature.AudioFeatures(sample_rate=22050,
frame_len=1024,
n_bands=2 * n_mfcc + 4,
n_audcc=n_mfcc,
half_fft=True,
hamming=True,
scale_to_int_range=True,
critical_band_fn=mel.melhtk_4k,
power_sqr_abs=True,
noise_level=1)
mfcc = af.audcc(frames)
mfcc_segments = audiofeature.segment(mfcc, n_segments)
return af, mfcc_segments, n_mfcc
def mfcc_features(frames, n_mfcc=16, n_segments=10):
af = audiofeature.AudioFeatures(
sample_rate=22050,
frame_len=1024,
n_bands=2 * n_mfcc + 4,
n_audcc=n_mfcc,
half_fft=True,
hamming=True,
scale_to_int_range=True,
critical_band_fn=mel.melhtk_4k,
power_sqr_abs=True,
noise_level=1,
)
mfcc = af.audcc(frames)
mfcc_segments = audiofeature.segment(mfcc, n_segments)
return af, mfcc_segments, n_mfcc
def mpc_w_clipping(frames, n_mpc=32, n_segments=10, use_sparse_warp=False):
AF = audiofeature.AudioFeatures(
22050,
1024,
n_bands=n_mpc,
scale_to_int_range=False,
critical_band_fn=mel.MelMasters.warp_dense,
power_sqr_abs=False,
use_sparse_warp=use_sparse_warp,
noise_level=1.0e-4,
) # 1/10000
audspec = AF.audspec(frames)
mpc = theano.tensor.log10(theano.tensor.switch(audspec > AF.noise_level, audspec, AF.noise_level))
mpc_segments = audiofeature.segment(mpc, n_segments)
return AF, mpc_segments, n_mpc
def mpc_w_clipping(frames, n_mpc=32, n_segments=10, use_sparse_warp=False):
AF = audiofeature.AudioFeatures(22050,
1024,
n_bands=n_mpc,
scale_to_int_range=False,
critical_band_fn=mel.MelMasters.warp_dense,
power_sqr_abs=False,
use_sparse_warp=use_sparse_warp,
noise_level=1.0e-4) #1/10000
audspec = AF.audspec(frames)
mpc = theano.tensor.log10(
theano.tensor.switch(audspec > AF.noise_level, audspec,
AF.noise_level))
mpc_segments = audiofeature.segment(mpc, n_segments)
return AF, mpc_segments, n_mpc
def mpc_w_max_approx_learnable(frames, n_mpc=32, n_segments=10, use_sparse_warp=False):
def log_warp_mat_fn(*args, **kwargs):
m = mel.MelMasters.warp_dense(*args, **kwargs)
return numpy.log10(m + 1.0e-12)
AF = audiofeature.AudioFeatures(
22050,
1024,
n_bands=n_mpc,
scale_to_int_range=False,
critical_band_fn=log_warp_mat_fn,
power_sqr_abs=False,
noise_level=1.0e-4,
) # 1/10000
powspec = AF.powspec(frames) # abs(fft(x))
logspec = theano.tensor.log10(powspec + AF.noise_level)
logspec3 = theano.tensor.DimShuffle(logspec.broadcastable, [0, "x", 1])(logspec)
sum3 = theano.tensor.add(AF.critical_band_warp_dense, logspec3)
feature = theano.tensor.max(sum3)
mpc_segments = audiofeature.segment(feature, n_segments)
return AF, mpc_segments, n_mpc
def mpc_w_max_approx(frames, n_mpc=32, n_segments=10, use_sparse_warp=False):
AF = audiofeature.AudioFeatures(
22050,
1024,
n_bands=n_mpc,
scale_to_int_range=False,
critical_band_fn=mel.MelMasters.warp_dense,
power_sqr_abs=False,
noise_level=1.0e-4,
) # 1/10000
powspec = AF.powspec(frames) # abs(fft(x))
logspec = theano.tensor.log10(powspec + AF.noise_level)
logspec3 = theano.tensor.DimShuffle(logspec.broadcastable, [0, "x", 1])(logspec)
# warp_mat dims: nfilts x nfft
warp_mat = AF.critical_band_fn(nfft=AF.frame_len / 2, fft_max_freq=AF.sample_rate / 2, nfilts=AF.n_bands)
log_warp_mat = numpy.log10(warp_mat + 1.0e-12)
sum3 = theano.tensor.add(log_warp_mat, logspec3)
feature = theano.tensor.max(sum3)
mpc_segments = audiofeature.segment(feature, n_segments)
return AF, mpc_segments, n_mpc
def mpc_w_max_approx(frames, n_mpc=32, n_segments=10, use_sparse_warp=False):
AF = audiofeature.AudioFeatures(22050,
1024,
n_bands=n_mpc,
scale_to_int_range=False,
critical_band_fn=mel.MelMasters.warp_dense,
power_sqr_abs=False,
noise_level=1.0e-4) #1/10000
powspec = AF.powspec(frames) #abs(fft(x))
logspec = theano.tensor.log10(powspec + AF.noise_level)
logspec3 = theano.tensor.DimShuffle(logspec.broadcastable,
[0, 'x', 1])(logspec)
# warp_mat dims: nfilts x nfft
warp_mat = AF.critical_band_fn(nfft=AF.frame_len / 2,
fft_max_freq=AF.sample_rate / 2,
nfilts=AF.n_bands)
log_warp_mat = numpy.log10(warp_mat + 1.0e-12)
sum3 = theano.tensor.add(log_warp_mat, logspec3)
feature = theano.tensor.max(sum3)
mpc_segments = audiofeature.segment(feature, n_segments)
return AF, mpc_segments, n_mpc
def mpc_w_max_approx_learnable(frames,
n_mpc=32,
n_segments=10,
use_sparse_warp=False):
def log_warp_mat_fn(*args, **kwargs):
m = mel.MelMasters.warp_dense(*args, **kwargs)
return numpy.log10(m + 1.0e-12)
AF = audiofeature.AudioFeatures(22050,
1024,
n_bands=n_mpc,
scale_to_int_range=False,
critical_band_fn=log_warp_mat_fn,
power_sqr_abs=False,
noise_level=1.0e-4) #1/10000
powspec = AF.powspec(frames) #abs(fft(x))
logspec = theano.tensor.log10(powspec + AF.noise_level)
logspec3 = theano.tensor.DimShuffle(logspec.broadcastable,
[0, 'x', 1])(logspec)
sum3 = theano.tensor.add(AF.critical_band_warp_dense, logspec3)
feature = theano.tensor.max(sum3)
mpc_segments = audiofeature.segment(feature, n_segments)
return AF, mpc_segments, n_mpc
