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']))
feat, energy = base.mfcc(sig, rate, self.conf)
if self.conf['include_energy'] == 'True':
feat = np.append(feat, energy[:, np.newaxis], 1)
if self.conf['dynamic'] == 'delta':
feat = base.delta(feat)
elif self.conf['dynamic'] == 'ddelta':
feat = base.ddelta(feat)
elif self.conf['dynamic'] != 'nodelta':
raise Exception('unknown dynamic type')
return feat
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
'''
feat, energy = base.mfcc(sig, rate, self.conf)
if self.conf['include_energy'] == 'True':
feat = np.append(feat, energy[:, np.newaxis], 1)
if self.conf['dynamic'] == 'delta':
feat = base.delta(feat)
elif self.conf['dynamic'] == 'ddelta':
feat = base.ddelta(feat)
elif self.conf['dynamic'] != 'nodelta':
raise Exception('unknown dynamic type')
#mean and variance normalize the features
if self.conf['mvn'] == 'True':
feat = (feat - feat.mean(0)) / feat.std(0)
return feat
def extractFeatures(input_signal):
"""extract features from the cleaned signal.
:param cleaned signal
:return: features list"""
# compute mfcc list
if len(input_signal) == 0:
print("cleaned signal is empty")
return input_signal
mfcc_list = np.array(mfcc(input_signal, samplerate=prm.params["sample_rate"].get(), winlen=0.032, winstep=0.016, numcep=30,
nfilt=55, nfft=2048, lowfreq=0, highfreq=6000, preemph=0.95, ceplifter=22, appendEnergy=True ))
extractor = LPCExtractor(prm.params["sample_rate"].get(), 32, 16, 30, 0.95)
lpcc = extractor.extract(input_signal)
pitch = extract_pitch(input_signal)
# Cepstral Mean Normalization @TODO: WHY IS THIS NOT HELPING??
if 0:
mean_mfcc = np.mean(mfcc_list.T, 1)
std_mfcc = np.std(mfcc_list.T, 1)
for i in range(len(mfcc_list)):
for j in range(len(mfcc_list[i])):
mfcc_list[i][j] = (mfcc_list[i][j]-mean_mfcc[j])/std_mfcc[j]
# print np.shape(mfcc_list[i-1]), np.shape(mean_mfcc), np.shape(std_mfcc)
N = 2
delta_list = delta(mfcc_list, N)
ddelta_list = delta(delta_list, N)
# do not keep first coeff (energy)
features_list = list()
for k in range(len(mfcc_list)):
# features_list += [np.hstack((mfcc_list[k][0:], lpcc[k][0:]))]
features_list += [mfcc_list[k][0:]]
# features_list += [lpcc[k][0:]]
# features_list += [np.hstack((mfcc_list[k][0:], delta_list[k][0:], ddelta_list[k][0:]))]
# print np.shape(mfcc_list), np.shape(features_list)
# dont return nan
# @TODO WHY DOES THIS HAPPEN?
for row in features_list:
for cell in row:
if cell != cell:
print "Cell is nan (see feature extraction):", str(cell)
return []
full_features_list = []
full_features_list = list(np.ravel(features_list))
full_features_list.extend([pitch]*30) # do we need to append this multiple times to ensure that the forest selects it?
return full_features_list
def mfcc_features(sig):
features = base.mfcc(sig,
samplerate=44100,
winlen=0.02,
winstep=0.01,
numcep=13,
nfilt=40)
# Mean Normalization for feature vectors.
mean_vector = np.mean(features, axis=0)
normalized = features - mean_vector
return normalized
liste_fe_echant = []
liste_classe = []
liste_dsp = []
matrice = []
for NomFichier in liste_nom_fichier_apprentissage:
(Fe, Echantillons) = scipy.io.wavfile.read("Signaux/" + NomFichier)
liste_fe_echant.append((Fe, Echantillons))
NumerClasse = Prefixe.index(NomFichier[0:2])
liste_classe.append(NumerClasse)
dsp = np.abs(np.fft.fft(Echantillons))
liste_dsp.append(dsp)
VecteurCoefficients = base.mfcc(Echantillons,
samplerate=Fe,
winlen=(len(Echantillons) / Fe),
winstep=(len(Echantillons) / Fe),
nfft=1024)
matrice.append(VecteurCoefficients[0])
# Affichage représentation temporelle et frequentielle (3 fichiers aléatoires)
nb1 = random.randint(0, len(liste_fe_echant))
nb2 = random.randint(0, len(liste_fe_echant))
nb3 = random.randint(0, len(liste_fe_echant))
plt.subplot(321)
plt.plot(liste_fe_echant[nb1][1], "r")
plt.title("Représentation temporelle de " + Prefixe[liste_classe[nb1]])
plt.subplot(322)
plt.plot(liste_dsp[nb1], "r")
plt.title("Représentation frequentielle de " + Prefixe[liste_classe[nb1]])
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
'''
feat, energy = base.mfcc(sig, rate, self.conf)
# write the wav to temporary location and invoke external pitch extractor.
# make sure 'reaper' is in your $PATH
tempdir = os.path.join('/tmp', str(os.getpid()))
if not os.path.isdir(tempdir):
os.makedirs(tempdir)
name = 'mix'
wav.write(os.path.join(tempdir, name + '.wav'), rate, np.int16(sig))
os.system('reaper -i ' + os.path.join(tempdir, name + '.wav') +
' -f ' + os.path.join(tempdir, name + '.txt -a -e 0.01'))
pitch = np.loadtxt(os.path.join(tempdir, name + '.txt'), skiprows=7)[:,
2]
pitch = np.pad(pitch, (0, max(0, feat.shape[0] - pitch.shape[0])),
'edge')
# linear interpolation in voiceless regions
voiceless = np.where(pitch == -1)[0]
jump = np.where((voiceless[1:] - voiceless[:-1]) > 1)[0]
segments = np.split(voiceless, jump + 1)
for seg in segments:
idx1 = seg[0] - 1
idx2 = seg[-1] + 1
val1 = -1
val2 = -1
if idx1 >= 0:
val1 = pitch[idx1]
if idx2 < pitch.size:
val2 = pitch[idx2]
if val1 == -1: #segment starts at utterence start
val1 = val2
if val2 == -1: # segment ends at utterance end
val2 = val1
if val1 == -1: #segment is the whole utterance => make up a value
val1 = 150
val2 = 150
#interpolate
pitch[seg] = (val2 - val1) * (np.array(seg) -
idx1) / float(idx2 - idx1) + val1
feat = np.append(feat, pitch[:feat.shape[0], np.newaxis], 1)
if self.conf['include_energy'] == 'True':
feat = np.append(feat, energy[:, np.newaxis], 1)
if self.conf['dynamic'] == 'delta':
feat = base.delta(feat)
elif self.conf['dynamic'] == 'ddelta':
feat = base.ddelta(feat)
elif self.conf['dynamic'] != 'nodelta':
raise Exception('unknown dynamic type')
#mean and variance normalize the features
if self.conf['mvn'] == 'True':
feat = (feat - feat.mean(0)) / (
feat.std(0) + 1e-20
) # features could be constant, e.g. voiceless speech
return feat
# 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.
# ==============================================================================
#!/usr/bin/env python
''' Example for sigproc.py '''
# pylint: skip-file
import scipy.io.wavfile as wav
from base import mfcc
from base import delta
from base import logfbank
if __name__ == '__main__':
(rate, sig) = wav.read("english.wav")
mfcc_feat = mfcc(sig, rate)
d_mfcc_feat = delta(mfcc_feat, 2)
fbank_feat = logfbank(sig, rate)
print(fbank_feat[1:3, :])
for f in batch1_fns:
# Get set of signals from 1 experiment with the highest value per channel
v1, v2, ev = read_ae_file2(f)
sig = []
for i in range(len(v1)):
sig.append(max_sig(v1[i], v2[i]))
sig = np.array(sig)
# jank code that converts raw signal to vector of mfcc
holder = []
for i in range(len(sig)):
holder.append(
base.mfcc(sig[i],
samplerate=rate,
winlen=window,
winstep=ratio,
lowfreq=300000,
highfreq=1800000))
X = []
for i in range(len(sig)):
X.append(holder[i][0])
'''
CLUSTER STATISTICS ROUTINE
'''
silh = np.array([]) # holder arrays
db_score = np.array([])
# Cluster and get stat
for i in range(min_cluster, max_cluster + 1):
kmeans = KMeans(n_clusters=i, n_init=100, tol=1e-6).fit(X)