def _dft_magnitude(self, signal):
"""Compute DFT and then its magnitude.
It is avoiding tflite incompatible ops.
Args:
signal: has dims [..., frame_size]
Returns:
magnitude_spectrogram: with dims [..., fft_size]
"""
real_spectrum = tf.matmul(signal, self.real_dft_tensor)
imag_spectrum = tf.matmul(signal, self.imag_dft_tensor)
magnitude_spectrum = tf.add(real_spectrum * real_spectrum,
imag_spectrum * imag_spectrum)
if self.magnitude_squared:
return magnitude_spectrum
else:
return tf.sqrt(magnitude_spectrum)
def call(self, inputs):
if self.use_tf:
# compute DCT with tf function
output = tf.signal.dct(inputs,
type=2,
n=self.num_features,
norm=None)
return output * self.norm
else:
# compute DCT direct matmul
return tf.matmul(inputs, self.dct)
def call(self, inputs):
# compute magnitude of fourier spectrum
fft_mag = super(MagnitudeRDFTmel, self).call(inputs)
# apply mel spectrum
if self.use_tf_fft:
mel_weight_matrix = tf.signal.linear_to_mel_weight_matrix(
num_mel_bins=self.num_mel_bins,
num_spectrogram_bins=self.feature_size,
sample_rate=self.sample_rate,
lower_edge_hertz=self.lower_edge_hertz,
upper_edge_hertz=self.upper_edge_hertz,
dtype=tf.float32)
else:
mel_weight_matrix = self.mel_weight_matrix
return tf.matmul(fft_mag, mel_weight_matrix)
def call(self, inputs):
# Multiply by mel weight matrix (num_spectrogram_bins, num_mel_bins)
return tf.matmul(inputs, self.mel_weight_matrix)
def call(self, inputs):
# compute DCT
return tf.matmul(inputs, self.dct)
def call(self, inputs):
# compute magnitude of fourier spectrum
fft_mag = super(MagnitudeRDFTmel, self).call(inputs)
# apply mel spectrum
return tf.matmul(fft_mag, self.mel_weight_matrix)
