def feature_extraction_lp_group_delay(y, fs=44100, statistics=True, lpgd_params=None, win_params=None):
eps = numpy.spacing(1)
nfft = lpgd_params['nfft']
lp_order = lpgd_params['lp_order']
y = y + eps
frames = segment_axis(y, win_params['win_length'], win_params['hop_length']);
print 'frames : ' + str(frames.shape)
a,e,k = lpc(frames, lp_order)
print 'a : ' + str(a.shape)
A = fft(a, nfft)
A = 1/A
phaseA = numpy.unwrap(numpy.angle(A))
print 'phaseA: ' + str(phaseA.shape)
phaseA = phaseA[:,0:nfft/2]
print 'phaseA: ' + str(phaseA.shape)
tauA = -1 * numpy.diff(phaseA)
print 'tauA' + str(tauA.shape)
# tau = numpy.concatenate((tauA, tauA[-1]))
# tau = tau
feature_matrix = tauA
feature_matrix = dct(feature_matrix, n=20)
print 'fm: ' + str(feature_matrix.shape)
# Collect into data structure
if statistics:
return {
'feat': feature_matrix,
'stat': {
'mean': numpy.mean(feature_matrix, axis=0),
'std': numpy.std(feature_matrix, axis=0),
'N': feature_matrix.shape[0],
'S1': numpy.sum(feature_matrix, axis=0),
'S2': numpy.sum(feature_matrix ** 2, axis=0),
}
}
else:
return {
'feat': feature_matrix}
def feature_extraction_lp_group_delay(y, fs=44100, statistics=True, lpgd_params=None, win_params=None):
eps = numpy.spacing(1)
nfft = lpgd_params['nfft']
lp_order = lpgd_params['lp_order']
y = y + eps
frames = segment_axis(y, win_params['win_length'], win_params['hop_length']);
print 'frames : ' + str(frames.shape)
a,e,k = lpc(frames, lp_order)
print 'a : ' + str(a.shape)
A = fft(a, nfft)
A = 1/A
phaseA = numpy.unwrap(numpy.angle(A))
print 'phaseA: ' + str(phaseA.shape)
phaseA = phaseA[:,0:nfft/2]
print 'phaseA: ' + str(phaseA.shape)
tauA = -1 * numpy.diff(phaseA)
print 'tauA' + str(tauA.shape)
# tau = numpy.concatenate((tauA, tauA[-1]))
# tau = tau
feature_matrix = tauA
feature_matrix = dct(feature_matrix, n=20)
print 'fm: ' + str(feature_matrix.shape)
# Collect into data structure
if statistics:
return {
'feat': feature_matrix,
'stat': {
'mean': numpy.mean(feature_matrix, axis=0),
'std': numpy.std(feature_matrix, axis=0),
'N': feature_matrix.shape[0],
'S1': numpy.sum(feature_matrix, axis=0),
'S2': numpy.sum(feature_matrix ** 2, axis=0),
}
}
else:
return {
'feat': feature_matrix}
def feature_extraction_gd(y, fs=44100, statistics=True, include_delta=True,
include_acceleration=True, lpgd_params=None, win_params=None, delta_params=None, acceleration_params=None):
eps = numpy.spacing(1)
nfft = lpgd_params['nfft']
lp_order = lpgd_params['lp_order']
y = y + eps
frames = segment_axis(y, win_params['win_length'], win_params['hop_length']);
#print 'frames : ' + str(frames.shape)
a,e,k = lpc(frames, lp_order)
#print 'a : ' + str(a.shape)
A = fft(a, nfft)
A = 1/A
phaseA = numpy.unwrap(numpy.angle(A))
#print 'phaseA: ' + str(phaseA.shape)
phaseA = phaseA[:,0:nfft/2]
#print 'phaseA: ' + str(phaseA.shape)
tauA = -1 * numpy.diff(phaseA)
#print 'tauA' + str(tauA.shape)
tauA = numpy.vstack((tauA, tauA[-1]))
# tau = tau
# print 'tauA' + str(tauA.shape)
feature_matrix = tauA
feature_matrix = dct(feature_matrix, n=20)
feature_matrix = feature_matrix.T
print feature_matrix.shape
if include_delta:
# Delta coefficients
feature_delta = librosa.feature.delta(feature_matrix, **delta_params)
# Add Delta Coefficients to feature matrix
feature_matrix = numpy.vstack((feature_matrix, feature_delta))
# print 'fm: ' + str(feature_matrix.shape)
if include_acceleration:
# Acceleration coefficients (aka delta)
feature_delta2 = librosa.feature.delta(feature_delta, order=2, **acceleration_params)
# Add Acceleration Coefficients to feature matrix
feature_matrix = numpy.vstack((feature_matrix, feature_delta2))
feature_matrix = feature_matrix.T
print 'fm: ' + str(feature_matrix.shape)
# Collect into data structure
if statistics:
return {
'tauA' : tauA,
'feat': feature_matrix,
'stat': {
'mean': numpy.mean(feature_matrix, axis=0),
'std': numpy.std(feature_matrix, axis=0),
'N': feature_matrix.shape[0],
'S1': numpy.sum(feature_matrix, axis=0),
'S2': numpy.sum(feature_matrix ** 2, axis=0),
}
}
else:
return {
'feat': feature_matrix}
def feature_extraction_gd(y,
fs=44100,
statistics=True,
include_delta=True,
include_acceleration=True,
lpgd_params=None,
win_params=None,
delta_params=None,
acceleration_params=None):
eps = numpy.spacing(1)
nfft = lpgd_params['nfft']
lp_order = lpgd_params['lp_order']
y = y + eps
frames = segment_axis(y, win_params['win_length'],
win_params['hop_length'])
#print 'frames : ' + str(frames.shape)
a, e, k = lpc(frames, lp_order)
#print 'a : ' + str(a.shape)
A = fft(a, nfft)
A = 1 / A
phaseA = numpy.unwrap(numpy.angle(A))
#print 'phaseA: ' + str(phaseA.shape)
phaseA = phaseA[:, 0:nfft / 2]
#print 'phaseA: ' + str(phaseA.shape)
tauA = -1 * numpy.diff(phaseA)
#print 'tauA' + str(tauA.shape)
tauA = numpy.vstack((tauA, tauA[-1]))
# tau = tau
# print 'tauA' + str(tauA.shape)
feature_matrix = tauA
feature_matrix = dct(feature_matrix, n=20)
feature_matrix = feature_matrix.T
print feature_matrix.shape
if include_delta:
# Delta coefficients
feature_delta = librosa.feature.delta(feature_matrix, **delta_params)
# Add Delta Coefficients to feature matrix
feature_matrix = numpy.vstack((feature_matrix, feature_delta))
# print 'fm: ' + str(feature_matrix.shape)
if include_acceleration:
# Acceleration coefficients (aka delta)
feature_delta2 = librosa.feature.delta(feature_delta,
order=2,
**acceleration_params)
# Add Acceleration Coefficients to feature matrix
feature_matrix = numpy.vstack((feature_matrix, feature_delta2))
feature_matrix = feature_matrix.T
print 'fm: ' + str(feature_matrix.shape)
# Collect into data structure
if statistics:
return {
'tauA': tauA,
'feat': feature_matrix,
'stat': {
'mean': numpy.mean(feature_matrix, axis=0),
'std': numpy.std(feature_matrix, axis=0),
'N': feature_matrix.shape[0],
'S1': numpy.sum(feature_matrix, axis=0),
'S2': numpy.sum(feature_matrix**2, axis=0),
}
}
else:
return {'feat': feature_matrix}