han16.py 文件源码

def compute_spectrograms(filename):
    out_rate = 22050

    frames, rate = librosa.load(filename, sr=out_rate, mono=True)
    if len(frames) < out_rate:
        # if less then 1 second - can't process
        raise Exception("Audio duration is too short")

    normalized_audio = _normalize(frames)
    melspectr = librosa.feature.melspectrogram(y=normalized_audio, sr=out_rate, n_mels=N_MEL_BANDS, fmax=out_rate/2)
    logmelspectr = librosa.logamplitude(melspectr**2, ref_power=1.0)

    # now going through spectrogram with the stride of the segment duration
    for start_idx in range(0, logmelspectr.shape[1] - SEGMENT_DUR + 1, SEGMENT_DUR):
        yield logmelspectr[:, start_idx:start_idx + SEGMENT_DUR]