无法使用libreplaygain.so和numpy获得合理的结果

Question

我一次又一次地尝试使用libreplaygain.so (ReplayGain是一种计算音频响度的算法。)从python传递来自音频文件的数据。这是libreplaygain的the header file。一般来说，我对ctypes和C语言都不太了解，所以我希望这可能是我愚蠢的问题，对其他人来说也很明显！下面是我使用的脚本：

import numpy as np
from scipy.io import wavfile
import ctypes
replaygain = ctypes.CDLL('libreplaygain.so')


def calculate_replaygain(samples, frame_rate=44100):
    """
    inspired from https://github.com/vontrapp/replaygain
    """
    replaygain.gain_init_analysis(frame_rate)
    block_size = 10000
    channel_count = samples.shape[1]
    i = 0
    samples = samples.astype(np.float64)

    while i * block_size < samples.shape[0]:
        channel_left = samples[i*block_size:(i+1)*block_size,0]
        channel_right = samples[i*block_size:(i+1)*block_size,1]

        samples_p_left = channel_left.ctypes.data_as(ctypes.POINTER(ctypes.c_double))
        samples_p_right = channel_right.ctypes.data_as(ctypes.POINTER(ctypes.c_double))

        replaygain.gain_analyze_samples(samples_p_left, samples_p_right, channel_left.shape[0], channel_count)
        i += 1

    return replaygain.gain_get_chapter()


if __name__ == '__main__':
    frame_rate, samples = wavfile.read('directions.wav')
    samples = samples.astype(np.float64) / 2**15
    gain = calculate_replaygain(samples, frame_rate=frame_rate)
    print "Recommended gain: %f dB" % gain
    gain = calculate_replaygain(np.random.random((441000, 2)) * 2 - 1, frame_rate=44100)
    print "Recommended gain: %f dB" % gain

脚本运行，但我无法获得与命令行工具replaygain相同的值。事实上，我总是得到80.0。你可以用任何声音文件替换'directions.wav‘...并将结果与命令replaygain <soundfile.wav>的结果进行比较。

Answer 1

gain_get_chapter()返回一个double，但是the ctypes docs say“默认假设函数返回的是C类型”。你应该像这样做

replaygain.gain_get_chapter.restype = ctypes.c_double

您还应该检查gain_init_analysis和gain_analyze_samples的返回值；如果这两个返回值都不是1，则说明出了问题。(它们实际上是int，所以您不需要在那里做任何其他事情。)