def test_tf_np_mel_tables(self):
frame_size = 400
num_mel_bins = 80
lower_edge_hertz = 125.0
upper_edge_hertz = 7600.0
sample_rate = 16000.0
fft_size = 2**int(math.ceil(math.log(frame_size) / math.log(2.0)))
# run numpy implementation
mel_np = mel_table.SpectrogramToMelMatrix(
num_mel_bins=num_mel_bins,
num_spectrogram_bins=fft_size // 2 + 1,
audio_sample_rate=sample_rate,
lower_edge_hertz=lower_edge_hertz,
upper_edge_hertz=upper_edge_hertz)
# run TF implementation
mel_tf = tf.signal.linear_to_mel_weight_matrix(
num_mel_bins=num_mel_bins,
num_spectrogram_bins=fft_size // 2 + 1,
sample_rate=sample_rate,
lower_edge_hertz=lower_edge_hertz,
upper_edge_hertz=upper_edge_hertz,
dtype=tf.float32)
# compare numpy vs TF implementations
self.assertAllClose(mel_np, mel_tf, rtol=1e-5, atol=1e-4)
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):
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)