preprocessing.py 文件源码

def calc_power_spectrogram(audio_data, samplerate, n_mels=128, n_fft=512, hop_length=160):
  """
  Calculate power spectrogram from the given raw audio data

  Args:
    audio_data: numpyarray of raw audio wave
    samplerate: the sample rate of the `audio_data`
    n_mels: the number of mels to generate
    n_fft: the window size of the fft
    hop_length: the hop length for the window

  Returns: the spectrogram in the form [time, n_mels]

  """
  spectrogram = librosa.feature.melspectrogram(audio_data, sr=samplerate, n_mels=n_mels, n_fft=n_fft, hop_length=hop_length)

  # convert to log scale (dB)
  log_spectrogram = librosa.logamplitude(spectrogram, ref_power=np.max)

  # normalize
  normalized_spectrogram = normalize(log_spectrogram)

  return normalized_spectrogram.T