def _normalize(x):
return dict(
x, **{
key:
features.window_normalization(x[key],
**config.get("kwargs", {}))
})
def extract_features(signals, sample_rates, feattype, spec_kwargs, melspec_kwargs, mfcc_kwargs, db_spec_kwargs, feat_scale_kwargs, window_norm_kwargs):
tf.debugging.assert_rank(signals, 2, message="Input signals for feature extraction must be batches of mono signals without channels, i.e. of shape [B, N] where B is batch size and N number of samples.")
tf.debugging.assert_equal(sample_rates, [sample_rates[0]], message="Different sample rates in a single batch not supported, all signals in the same batch should have the same sample rate.")
#TODO batches with different sample rates (probably not worth the effort)
sample_rate = sample_rates[0]
X = audio_features.spectrograms(signals, sample_rate, **spec_kwargs)
tf.debugging.assert_all_finite(X, "spectrogram failed")
if feattype in ("melspectrogram", "logmelspectrogram", "mfcc"):
X = audio_features.melspectrograms(X, sample_rate=sample_rate, **melspec_kwargs)
tf.debugging.assert_all_finite(X, "melspectrogram failed")
if feattype in ("logmelspectrogram", "mfcc"):
X = tf.math.log(X + 1e-6)
tf.debugging.assert_all_finite(X, "logmelspectrogram failed")
if feattype == "mfcc":
coef_begin = mfcc_kwargs.get("coef_begin", 1)
coef_end = mfcc_kwargs.get("coef_end", 13)
mfccs = tf.signal.mfccs_from_log_mel_spectrograms(X)
X = mfccs[..., coef_begin:coef_end]
tf.debugging.assert_all_finite(X, "mfcc failed")
elif feattype in ("db_spectrogram",):
X = audio_features.power_to_db(X, **db_spec_kwargs)
tf.debugging.assert_all_finite(X, "db_spectrogram failed")
if feat_scale_kwargs:
X = features.feature_scaling(X, **feat_scale_kwargs)
tf.debugging.assert_all_finite(X, "feature scaling failed")
if window_norm_kwargs:
X = features.window_normalization(X, **window_norm_kwargs)
tf.debugging.assert_all_finite(X, "window normalization failed")
return X
def test_window_normalization(self):
rank = 3
for _ in range(50):
delta = np.random.uniform(1, 1e3)
x = np.random.uniform(-delta,
delta,
size=np.random.randint(1, 20, size=rank))
axis = 1
assert axis == 1, "TODO: test windows on arbitrary axis"
for window_len in [-1] + list(range(2, x.shape[0] + 1)):
for normalize_variance in (True, False):
kwargs = dict(axis=axis,
window_len=window_len,
normalize_variance=normalize_variance)
y = features.window_normalization(x, **kwargs).numpy()
assert not np.isnan(y).any()
assert y.shape == x.shape