def signals2logmel(signals, sample_rate, num_mel_bins):
signals, sample_rate = signals[:,::3], sample_rate // 3
flen = audio.ms_to_frames(sample_rate, 25)
fstep = audio.ms_to_frames(sample_rate, 10)
S = tf.math.square(tf.math.abs(tf.signal.stft(signals, flen, fstep, fft_length=512)))
# S = audio.spectrograms(signals, sample_rate)
S = audio.linear_to_mel(S, sample_rate, num_mel_bins=num_mel_bins, fmax=tf.cast(sample_rate, tf.float32))
S = tf.math.log(1e-6 + S)
S = cmvn(S, axis=1)
return S
def test_spectrograms(self):
for path in audiofiles:
s, r = audio.read_wav(path)
for len_ms in range(20, 101, 20):
for n_fft in (256, 512, 1024, 2048):
if n_fft < audio.ms_to_frames(r, len_ms):
continue
step_ms = len_ms // 2
powspec = audio.spectrograms(np.expand_dims(s, 0), r,
frame_length_ms=len_ms,
frame_step_ms=step_ms,
fft_length=n_fft)[0]
assert not np.isnan(powspec.numpy()).any()
assert powspec.shape[0] == s.shape[0] // audio.ms_to_frames(r, step_ms) - 1
assert powspec.shape[1] == n_fft // 2 + 1
def test_ms_to_frames(self):
for sr in range(1000, 60000, 1000):
for ms in range(1, 5000, 100):
nframes = (sr // 1000) * ms
assert audio.ms_to_frames(sr, ms).numpy() == nframes