preproc.py 文件源码

def mfcc_features(filename):
  """Preprocessing per CTC paper.

  (These are not the simpler linear spectrogram features alone as in Deep
  Speech).

  Properties:
  - 10ms frames with 5ms overlap
  - 12 MFCCs with 26 filter banks
  - replace first MFCC with energy (TODO: log-energy)
  - add first-order derivatives for all of the above
  - total: 26 coefficients
  """
  d, sr = librosa.load(filename)

  frame_length_seconds = 0.010
  frame_overlap_seconds = 0.005

  mfccs = librosa.feature.mfcc(d, sr, n_mfcc=1+12, n_fft=int(frame_overlap_seconds*sr), hop_length=int(frame_overlap_seconds*sr))

  # energy (TODO: log?)
  energy = librosa.feature.rmse(d, n_fft=int(frame_overlap_seconds*sr), hop_length=int(frame_overlap_seconds*sr))

  mfccs[0] = energy # replace first MFCC with energy, per convention

  deltas = librosa.feature.delta(mfccs, order=1)
  mfccs_plus_deltas = np.vstack([mfccs, deltas])

  coeffs = sklearn.preprocessing.scale(mfccs_plus_deltas, axis=1)

  return coeffs