def test_mel(sample_rate, n_freq, n_mels, f_min, f_max_ratio, htk, norm):
f_max = int(f_max_ratio * (sample_rate // 2))
mel_fb = KPB.filterbank_mel(
sample_rate=sample_rate,
n_freq=n_freq,
n_mels=n_mels,
f_min=f_min,
f_max=f_max,
htk=htk,
norm=norm,
)
mel_fb = mel_fb.numpy()
mel_fb_ref = librosa.filters.mel(
sr=sample_rate,
n_fft=(n_freq - 1) * 2,
n_mels=n_mels,
fmin=f_min,
fmax=f_max,
htk=htk,
norm=norm,
).T
assert mel_fb.dtype == K.floatx()
assert mel_fb.shape == (n_freq, n_mels)
np.testing.assert_allclose(mel_fb_ref, mel_fb)
def build(self, input_shape):
if self.image_data_format == 'channels_first':
self.n_ch = input_shape[1]
self.n_freq = input_shape[2]
self.n_time = input_shape[3]
else:
self.n_ch = input_shape[3]
self.n_freq = input_shape[1]
self.n_time = input_shape[2]
if self.init == 'mel':
self.filterbank = K.variable(backend.filterbank_mel(sr=self.sr,
n_freq=self.n_freq,
n_mels=self.n_fbs,
fmin=self.fmin,
fmax=self.fmax).transpose(),
dtype=K.floatx())
elif self.init == 'log':
self.filterbank = K.variable(backend.filterbank_log(sr=self.sr,
n_freq=self.n_freq,
n_bins=self.n_fbs,
bins_per_octave=self.bins_per_octave,
fmin=self.fmin).transpose(),
dtype=K.floatx())
if self.trainable_fb:
self.trainable_weights.append(self.filterbank)
else:
self.non_trainable_weights.append(self.filterbank)
super(Filterbank, self).build(input_shape)
self.built = True
def test_filterbank_mel():
"""test for backend.filterbank_mel"""
fbmel_ref = np.load(
os.path.join(os.path.dirname(__file__), 'fbmel_8000_512.npy'))
fbmel = KPB.filterbank_mel(sr=8000, n_freq=512)
assert fbmel.shape == fbmel_ref.shape
assert np.allclose(fbmel, fbmel_ref, atol=TOL)