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, energy = base.logfbank_scipy(sig, rate, self.conf)
else:
feat, energy = base.logfbank(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 testit(X, Y, batch_size, MODEL):
n_batch = X.shape[0] / batch_size
y_hat = []
accu = []
for batch in xrange(n_batch):
if (isinstance(MODEL, BULBUL)):
x_batch = asarray([
base.logfbank(x,
samplerate=22050,
winlen=0.046,
winstep=0.010,
nfilt=80,
nfft=1024,
lowfreq=10,
highfreq=11000).T
for x in X[batch * batch_size:(batch + 1) *
batch_size].astype('float32')
]).astype('float32')
else:
x_batch = X[batch * batch_size:(batch + 1) *
batch_size].astype('float32')
y_batch = Y[batch * batch_size:(batch + 1) *
batch_size].astype('int32')
accu.append(MODEL.test(x_batch, y_batch))
y_hat.append(MODEL.predict(x_batch))
accuracy = array(accu).mean()
print 'ACCURACY ', accuracy
y_hat = concatenate(y_hat, axis=0)
y_hat = vstack([y_hat.reshape((1, -1)), 1 - y_hat.reshape((1, -1))]).T
return roc_auc_score(Y[:shape(y_hat)[0]], y_hat.argmin(axis=1)), accuracy
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.logfbank(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 trainit(X_train, Y_train, X_test, Y_test, batch_size, n_epochs, MODEL,
index_0, index_1, l_r):
n_batch = X_train.shape[0] / batch_size
train_error = []
test_error = []
filters = []
representation = []
accuracy = []
for n_epoch in xrange(n_epochs):
if (n_epoch < 10 or n_epoch == (n_epochs - 1)):
filters.append(MODEL.get_filters())
for batch in xrange(n_batch):
index_0_batch = random.sample(index_0, batch_size / 2)
index_1_batch = random.sample(index_1, batch_size / 2)
index_batch = concatenate([index_0_batch, index_1_batch])
x_batch = X_train[index_batch]
y_batch = Y_train[index_batch]
if (isinstance(MODEL, BULBUL)):
x_batch = asarray([
base.logfbank(x,
samplerate=22050,
winlen=0.046,
winstep=0.010,
nfilt=80,
nfft=1024,
lowfreq=10,
highfreq=11000).T for x in x_batch
]).astype('float32')
train_error.append(
MODEL.train(x_batch, y_batch, l_r.astype('float32')))
if batch % 40 == 0:
print 'batch n_', batch, 'out of', n_batch, ': ', 'training error=', train_error[
-1]
auc, accu = testit(X_test, Y_test, batch_size, MODEL)
test_error.append(auc)
accuracy.append(accu)
print "epoch n_", n_epoch, ' AUC= ', test_error[-1]
index_0 = find(Y_test == 0)
index_1 = find(Y_test == 1)
for i in xrange(5):
index_batch = concatenate([
index_0[i * batch_size / 2:(i + 1) * batch_size / 2],
index_1[i * batch_size / 2:(i + 1) * batch_size / 2]
])
x_batch = X_train[index_batch]
y_batch = Y_train[index_batch]
if (isinstance(MODEL, BULBUL)):
x_batch = asarray([
base.logfbank(x,
samplerate=22050,
winlen=0.046,
winstep=0.010,
nfilt=80,
nfft=1024,
lowfreq=10,
highfreq=11000).T for x in x_batch
]).astype('float32')
representation.append([MODEL.get_repr(x_batch), y_batch])
f = open(name, 'wb')
cPickle.dump([train_error, test_error, accuracy, filters, representation],
f)
f.close()
# 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, :])