process.py 文件源码

def load_wav_file(fname, smprate=16000):
    '''
    load a WAV file, then return a numpy float32 vector.
    Resample if needed.

    The returned array will always have lenght of multiples of FFT_SIZE
    to ease preprocessing, this is done via zero padding at the end.

    '''
    smprate_real, data = wavfile.read(fname)
    if smprate_real == smprate:
        data = data.astype(FLOATX)
    elif (smprate_real % smprate) == 0:
        # integer factor downsample
        smpfactor = smprate_real // smprate
        data = np.pad(
            data, [(0, (-len(data)) % smpfactor)], mode='constant')
        data = np.reshape(data, [len(data)//smpfactor, smpfactor])
        data = np.mean(data.astype(FLOATX), axis=1)
    else:
        newlen = int(ceil(len(data) * (smprate / smprate_real)))
        # FIXME this resample is very slow on prime length
        data = scipy.signal.resample(data, newlen).astype(FLOATX)
    return data