OLA算法代码示例

import matplotlib.pyplot as plt
import numpy as np
import wave
import librosa

# w:要改变的声音 N:分割长度 S:分割步长 Ss:合成步长
def OLA_Curse(w:np.ndarray[np.float64],N:int,S:int,Ss:int):
    # 分割
    print()
    Cut_Buffer = [] # 分割缓冲区
    i = 0
    while i < w.size:
        print('\r',i,w.size,end='')
        if i+N > w.size:
            Cut_Buffer.append(w[i:]) # 末尾
        else:
            Cut_Buffer.append(w[i:i+N]) # 步长为S，帧长度为N
        i+=S # 步长
    #print(Cut_Buffer)
    
    # 合成(速度很慢)
    # 返回的音轨
    track = np.zeros(len(Cut_Buffer)*Ss)
    i = 0
    print('')
    n = 0
    for frame in Cut_Buffer:
        print('\r',n,len(Cut_Buffer),end='')
        # 确保音轨足够长
        '''
        if i+frame.size > track.size:
            track = np.append(track,np.zeros(i+frame.size-track.size))
        '''
        # 重叠相加
        for j in frame:
            track[i] += j
            i+=1
        i -= N-Ss # 回滚到下一帧起始位置
        n+=1
    
    #print(track)
    print('')
    return track


if __name__ == '__main__':
    
    print('READY')
    track = librosa.load('test1.wav',sr=64000)[0]
    track1 = track.astype(np.float64)
    
    print('Converting')
    track_tmp = OLA_Curse(track,6400,3200,6400)
    
    print(track.size,track_tmp.size)
    
    
    bdat = (track_tmp*(32000.0/track_tmp.max())).astype(np.int16).tobytes()
    f =wave.open('test.wav','wb')
    f.setnchannels(1)
    f.setsampwidth(2)
    f.setframerate(64000)
    f.writeframes(\
        bdat\
    )
    f.close()

实践证明，噪声有点大，在cos函数构成的乐曲中有明显的噗噗声。