def feat_pipeline(vec, freq):
feats = base.compute_features(vec, freq, 1.0)
voice = Vector(compute_vad_energy(
vad_opts, feats)) # Use origin mfcc to computed
delta_feats = compute_deltas(delta_opts, feats)
sliding_feats = Matrix(delta_feats.num_rows, delta_feats.num_cols)
sliding_window_cmn(sliding_opts, delta_feats, sliding_feats)
if not voice.sum():
LOG.warning('No features were judged as voiced for utterance')
return False
dim = int(voice.sum())
voice_feats = Matrix(dim, delta_feats.num_cols)
feats = kaldi_Matrix(sliding_feats)
index = 0
for i, sub_vec in enumerate(feats):
if voice[i] != 0 and voice[i] == 1:
voice_feats.row(index).copy_row_from_mat_(feats, i)
index += 1
LOG.debug('Feats extract successed')
return voice_feats
def load_features(self, feat_rxspecifier, vad_rxspecifier):
feats = kio.read_matrix(feat_rxspecifier)
vad_labels = kio.read_vector(vad_rxspecifier)
feats = featfuncs.compute_deltas(self.delta_opts, feats)
featfuncs.sliding_window_cmn(self.cmn_opts, feats, feats)
feats = feats.numpy()[vad_labels.numpy().astype(bool), :]
return feats
def feat_pipeline(wav):
feats = base.compute_features(wav.data()[0], wav.samp_freq, 1.0)
cmvn = Cmvn(base.dim())
cmvn.accumulate(feats)
cmvn.apply(feats)
return compute_deltas(opts, feats)
def feat_pipeline(wav):
feats = base.compute_features(wav.data()[0], wav.samp_freq, 1.0)
return compute_deltas(opts, feats)