Librosa音频处理（二）

对音频信号的处理可以通过 librosa.ifgram 方法获取 stft 短时傅立叶变换的矩阵，对该矩阵进行修改搬移，再进行 istft 逆转换获得处理后的音频信号。

y, sr = librosa.load(path)
frequencies, D = librosa.ifgram(y, sr=sr)
'''
     中间对D进行处理就行了
'''
y = librosa.istft(D)

D为stft变换的矩阵，x 轴为时间序列，y轴为频率序列坐标对应frequencies，值为幅度。

由于D类型为numpy.ndarray，所以我们很方便就可以通过numpy库对矩阵处理。

 

 

1. 回音

   D = np.repeat(D, 2, axis=1)

2. 间断

   D[:,::2] = 0

    

3. 音色

   D = np.roll(D, 50, axis=0)

4. 压缩频率

   def _pool(D, poolsize):
       x = D.shape[1] // poolsize
       restsize = D.shape[1] % poolsize
       if restsize > 0:
           x += 1
           rightlist = np.zeros([ D.shape[0], poolsize-restsize])
           D = np.c_[D, rightlist]
       D = D.reshape( (-1, poolsize) )
       D = D.sum(axis=1).reshape(-1,x)
       return D
   
   def rewardshape(D, shape):
       x = shape[0] - D.shape[0]
       y = shape[1] - D.shape[1]
       if x > 0:
           bottomlist = np.zeros([x, D.shape[1]])
           D = np.r_[D, bottomlist]
       if y > 0:
           rightlist = np.zeros([ D.shape[0], y])
           D = np.c_[D, rightlist]
       return D
   
   def pool(D, size=(3,3), shapeed=False):
       _shape = D.shape
       if size[1] > 1:
           D = _pool(D, size[1])
       if size[0] > 1:
           D = _pool(D.T, size[0]).T
       if shapeed:
           D = rewardshape(D, _shape)
       return D