def _extractFeatures_(self):
print('Extraction des features ...')
mfcc_feat = []
filter_feat = []
spectrum = []
for i in range(len(self.rate)):
mfcc_feat.append(
mfcc(self.data[i], self.rate[i], self.winlen, self.winstep,
self.numcep, self.nfilt, self.nfft, self.lowfreq,
self.highfreq))
filterFeat, energy, spectrum0 = fbank(self.data[i], self.rate[i],
self.winlen, self.winstep,
self.nfilt, self.nfft,
self.lowfreq, self.highfreq)
filterFeat = 10 * np.log10(filterFeat)
#filterFeat -= (numpy.mean(filterFeat, axis=0) + 1e-8)
#filterFeat /=numpy.std(filterFeat, axis=0)
spectrum0 = 10 * np.log10(spectrum0)
#spectrum0 -= (numpy.mean(spectrum0, axis=0) + 1e-8)
#spectrum0 /=numpy.std(spectrum0, axis=0)
filter_feat.append(filterFeat)
spectrum.append(spectrum0)
self.features = {
'mfcc': mfcc_feat,
'filter': filter_feat,
'spectrum': spectrum
}
print('Extraction des features terminée ...')
def comp_feat(self, sig, rate):
"""
compute the features
Args:
sig: the audio signal as a 1-D numpy array
rate: the sampling rate
Returns:
the features as a [seq_length x feature_dim] numpy array
"""
# snip the edges
sig = snip(sig, rate, float(self.conf['winlen']),
float(self.conf['winstep']))
if 'scipy' in self.conf and self.conf['scipy'] == 'True':
feat = base.angspec_scipy(sig, rate, self.conf)
else:
feat = base.angspec(sig, rate, self.conf)
if self.conf['include_energy'] == 'True':
if 'scipy' in self.conf and self.conf['scipy'] == 'True':
_, energy = base.fbank_scipy(sig, rate, self.conf)
else:
_, energy = base.fbank(sig, rate, self.conf)
feat = np.append(feat, energy[:, np.newaxis], 1)
return feat