def call(self, audio_data, sample_rate=None):
"""
Caculate fbank features of audio data.
:param audio_data: the audio signal from which to compute spectrum. Should be an (1, N) tensor.
:param sample_rate: [option]the samplerate of the signal we working with, default is 16kHz.
:return: A float tensor of size (num_channels, num_frames, num_frequencies) containing
fbank features of every frame in speech.
"""
p = self.config
with tf.name_scope('fbank'):
if sample_rate == None:
sample_rate = tf.constant(p.sample_rate, dtype=tf.int32)
if p.upper_frequency_limit <= 0:
p.upper_frequency_limit = p.sample_rate / 2.0 + p.upper_frequency_limit
elif (p.upper_frequency_limit <= p.lower_frequency_limit) or (
p.upper_frequency_limit > p.sample_rate / 2.0):
p.upper_frequency_limit = p.sample_rate / 2.0
assert_op = tf.assert_equal(tf.constant(p.sample_rate),
tf.cast(sample_rate, dtype=tf.int32))
with tf.control_dependencies([assert_op]):
spectrum = self.spect(audio_data, sample_rate)
spectrum = tf.expand_dims(spectrum, 0)
fbank = py_x_ops.fbank(
spectrum,
sample_rate,
upper_frequency_limit=p.upper_frequency_limit,
lower_frequency_limit=p.lower_frequency_limit,
filterbank_channel_count=p.filterbank_channel_count)
return fbank
def test_fbank(self):
""" test fbank op"""
with self.session():
data = np.arange(513)
spectrogram = tf.constant(data[None, None, :], dtype=tf.float32)
sample_rate = tf.constant(22050, tf.int32)
output = py_x_ops.fbank(spectrogram,
sample_rate,
filterbank_channel_count=20)
output_true = np.array([
1.887894,
2.2693727,
2.576507,
2.8156495,
3.036504,
3.2296343,
3.4274294,
3.5987632,
3.771217,
3.937401,
4.0988584,
4.2570987,
4.4110703,
4.563661,
4.7140336,
4.8626432,
5.009346,
5.1539173,
5.2992935,
5.442024,
])
self.assertEqual(tf.rank(output).eval(), 3)
self.assertEqual(output.shape, (1, 1, 20))
self.assertAllClose(output.eval(), output_true[None, None, :])
def fbank_feat(powspec,
sr=8000,
feature_size=40,
nfft=512,
lowfreq=0,
highfreq=None):
''' powspec: [audio_channels, spectrogram_length, spectrogram_feat_dim]
return : [auido_chnnels, nframe, nfbank]
'''
del nfft
true_fn = lambda: tf.expand_dims(powspec, 0)
false_fn = lambda: powspec
powspec = tf.cond(tf.equal(tf.rank(powspec), 2), true_fn, false_fn)
feat = py_x_ops.fbank(
powspec,
sr,
filterbank_channel_count=feature_size,
lower_frequency_limit=lowfreq,
upper_frequency_limit=highfreq,
)
return feat
