基于小波变换和动态验证的基音检测算法

Pitch detection method based on wavelet transform and dynamic verification

针对基音周期检测中容易出现的半周期和倍周期错误,综合考虑了常用的小波变换和短时自相关方法的优缺点,以及相邻基音周期长度的渐变性,提出了把两者相结合的基音周期检测算法。对语音信号进行清浊音检测和前置带通滤波,利用小波变换方法进行初步检测,对基音周期变化过大的情况使用自相关方法进行验证。实验结果表明,该方法在不同信噪比下的基音周期检测准确率都明显高于普通的小波变换检测方法。同时,该方法还有助于通过人工方式快速修正基音周期。

Aiming at the half and double frequency error often appearing in pitch detections,advantages and disadvantages of wavelet transform and short time autocorrelation function are analyzed,gradual change of the neighbor pitches is also considered,a pitch detection method based on wavelet transform and dynamic verification is proposed.First,voiced/unvoiced classification and band-pass pre-filter is used on the speech signal.Then,wavelet transform is used for basic pitch detection.Finally,autocorrelation function is used to determine the accurate pitches.Experimental results show that the proposed method performs much better than the normal wavelet transform method in the cases of both clean and noise-degraded speech and it is helpful in manual modification of the pitches.