def build(self, input_shape):
# output size of DFT
feature_size = self._compute_fft_size(int(input_shape[-1])) // 2 + 1
self.feature_size = feature_size
fft_mel_size = None
if not self.use_tf_fft:
# precompute mel matrix using np
self.mel_weight_matrix = mel_table.SpectrogramToMelMatrix(
num_mel_bins=self.num_mel_bins,
num_spectrogram_bins=feature_size,
audio_sample_rate=self.sample_rate,
lower_edge_hertz=self.lower_edge_hertz,
upper_edge_hertz=self.upper_edge_hertz)
if self.mel_non_zero_only:
fft_mel_size = self._get_non_zero_mel_size()
self.mel_weight_matrix = self.mel_weight_matrix[:
fft_mel_size, :]
self.mel_weight_matrix = tf.constant(self.mel_weight_matrix,
dtype=tf.float32)
super(MagnitudeRDFTmel, self).build(input_shape, fft_mel_size)
def build(self, input_shape, fft_mel_size=None):
super(MagnitudeRDFT, self).build(input_shape)
frame_size = int(input_shape[-1])
self.fft_size = self._compute_fft_size(frame_size)
if (self.use_tf_fft and fft_mel_size):
raise ValueError('TF FFT(True) is not compatible with fft_mel_size')
if not self.use_tf_fft:
# it is a real DFT with cos and sin functions only
# for real and imaginary components accordingly:
dft_real = np.asarray(
np.cos(2.0 * np.pi *
np.outer(np.arange(self.fft_size), np.arange(self.fft_size)) /
self.fft_size),
dtype=np.float32)
dft_imag = np.asarray(
-np.sin(2.0 * np.pi *
np.outer(np.arange(self.fft_size), np.arange(self.fft_size)) /
self.fft_size),
dtype=np.float32)
if fft_mel_size is None:
dft_real_half = dft_real[:self.fft_size // 2 + 1, :]
dft_imag_half = dft_imag[:self.fft_size // 2 + 1, :]
else:
dft_real_half = dft_real[:fft_mel_size, :]
dft_imag_half = dft_imag[:fft_mel_size, :]
dft_real = dft_real_half.transpose()
dft_imag = dft_imag_half.transpose()
# extract only array with size of input signal, so that
# there will be no need to do padding of input signal (it is not FFT)
# and there will be no multiplications with padded zeros
self.real_dft_tensor = tf.constant(dft_real[:frame_size, :])
self.imag_dft_tensor = tf.constant(dft_imag[:frame_size, :])
def build(self, input_shape):
super(MelSpectrogram, self).build(input_shape)
feature_size = int(input_shape[-1])
if self.use_tf:
# precompute mel matrix using tf
self.mel_weight_matrix = tf.signal.linear_to_mel_weight_matrix(
num_mel_bins=self.num_mel_bins,
num_spectrogram_bins=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:
# precompute mel matrix using np
self.mel_weight_matrix = tf.constant(
mel_table.SpectrogramToMelMatrix(
num_mel_bins=self.num_mel_bins,
num_spectrogram_bins=feature_size,
audio_sample_rate=self.sample_rate,
lower_edge_hertz=self.lower_edge_hertz,
upper_edge_hertz=self.upper_edge_hertz),
dtype=tf.float32)