def batch_extract_features(x):
with tf.device("GPU"):
signals, rates = x["signal"], x["sample_rate"]
logmelspecs = logmelspectrograms(signals, rates[0])
logmelspecs_smn = cmvn(logmelspecs)
mfccs = tf.signal.mfccs_from_log_mel_spectrograms(logmelspecs)
mfccs = mfccs[..., 1:21]
mfccs_cmvn = cmvn(mfccs)
return dict(x, logmelspec=logmelspecs_smn, mfcc=mfccs_cmvn)
def convertBrowserFFT(spec, sample_rate, num_mel_bins):
S = audio.db_to_power(spec)
# S = tf.math.abs(tf.signal.stft(signals, 400, 160, 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/2, tf.float32))
S = tf.math.log(1e-6 + S)
S = cmvn(S, axis=1)
return S
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_cmvn(self):
rank = 3
for delta_magnitude in range(2, 7):
for _ in range(200):
delta = np.random.uniform(1, 10**delta_magnitude)
x = np.random.uniform(-delta,
delta,
size=np.random.randint(1, 20, size=rank))
for axis in range(rank):
y_m = features.cmvn(x, axis=axis,
normalize_variance=False).numpy()
assert not np.isnan(y_m).any()
assert y_m.shape == x.shape
assert np.abs(y_m.mean(axis=axis)).max() < 1
y_mv = features.cmvn(x, axis=axis,
normalize_variance=True).numpy()
assert not np.isnan(y_mv).any()
assert y_mv.shape == x.shape
assert np.abs(y_mv.mean(axis=axis)).max() < 0.1
assert y_mv.var(axis=axis).max() < 10