def test_delta_delta(self):
''' test add delta detlas '''
#pylint: disable=invalid-name
p = tffeat.speech_params(sr=self.sr_true,
bins=40,
cmvn=False,
audio_desired_samples=1000,
add_delta_deltas=False)
with self.session():
wavfile = tf.constant(self.wavpath)
audio, sample_rate = tffeat.read_wav(wavfile, self.hp)
del sample_rate
feature = tffeat.compute_mel_filterbank_features(
audio,
sample_rate=p.audio_sample_rate,
preemphasis=p.audio_preemphasis,
frame_length=p.audio_frame_length,
frame_step=p.audio_frame_step,
lower_edge_hertz=p.audio_lower_edge_hertz,
upper_edge_hertz=p.audio_upper_edge_hertz,
num_mel_bins=p.audio_num_mel_bins,
apply_mask=False)
feature = tffeat.delta_delta(feature, order=2)
self.assertEqual(feature.eval().shape, (11, 40, 3))
def test_tf_spec(self):
''' test tensorflow spec feature interface '''
speech_feature.extract_feature((self.wavfile),
winlen=self.winlen,
winstep=self.winstep,
sr=self.sr,
feature_size=self.feature_size,
feature_name='spec')
feat = np.load(self.featfile)
self.assertEqual(feat.shape, (425, 129, 1))
with self.session(use_gpu=False):
feat = speech_ops.delta_delta(feat, 2)
self.assertEqual(feat.eval().shape, (425, 129, 3))
def test_tf_feat(self):
''' test tensorflow fbank feature interface '''
speech_feature.extract_filterbank((self.wavfile),
winlen=self.winlen,
winstep=self.winstep,
sr=self.sr,
feature_size=self.feature_size)
feat = np.load(self.featfile)
self.assertEqual(feat.shape, (425, 40, 1))
with self.session():
feat = speech_ops.delta_delta(feat, 2)
self.assertEqual(feat.eval().shape, (425, 40, 3))
def add_delta_delta(feat, feat_size, order=2):
''' add delta detla '''
graph = tf.Graph()
#pylint: disable=not-context-manager
with graph.as_default():
fbank = tf.placeholder(dtype=tf.float32,
shape=[None, feat_size, 1],
name='fbank')
feat_with_delta_delta = speech_ops.delta_delta(fbank, order=order)
sess = tf.Session(graph=graph)
feat = sess.run(feat_with_delta_delta, feed_dict={'fbank:0': feat})
sess.close()
return feat
def test_tf_fbank(self):
''' test tensorflow fbank feature interface '''
speech_feature.extract_feature((self.wavfile),
winlen=self.winlen,
winstep=self.winstep,
sr=self.sr,
feature_size=self.feature_size,
feature_name='fbank')
feat = np.load(self.featfile)
logging.info(f"feat : {feat}")
self.assertEqual(feat.shape, (425, 40, 1))
with self.session(use_gpu=False):
feat = speech_ops.delta_delta(feat, 2)
self.assertEqual(feat.eval().shape, (425, 40, 3))
def add_delta_delta(feat, feat_size, order=2):
''' add delta detla '''
feat_name = 'delta_delta'
graph = None
# get session
if feat_name not in _global_sess:
graph = tf.Graph()
#pylint: disable=not-context-manager
with graph.as_default():
fbank = tf.placeholder(
dtype=tf.float32, shape=[None, feat_size, 1], name='fbank')
feat_with_delta_delta = speech_ops.delta_delta(fbank, order=order)
feat_with_delta_delta = tf.identity(feat_with_delta_delta, name=feat_name)
sess = _get_session(feat_name, graph)
feat = sess.run(
_get_out_tensor_name(feat_name, 0), feed_dict={'fbank:0': feat})
return feat
