Abstract: Pitch detection is one of the extremely important steps in speech signal processing, but the detection result usually has errors because pitch is influenced by vocal tract characteristics and noise. The excitation signal and vocal tract characteristics in voiced speech are separated by using homomorphism deconvolution, so as to reduce the influence of vocal tract characteristics and improve the pitch detection accuracy. Then autocorrelation is used to detect pitch in excitation signal. The feasibility is verified by theoretical model. The experiment results of real speech signal indicate that the vocal tract characteristics and noise have rarely influence on pitch detection using this method.

Key words: homomorphism deconvolution, pitch detection, speech, frequency invariant linear filter

摘要： 语音处理中基音检测是极为重要的环节之一，然而浊音中的基音往往会受到声道特性和噪声的影响而导致检测结果的误差。利用同态解卷处理，将浊音中的激励信号和声道特性进行分离，然后再在激励信号中利用自相关检测基音，可以减小声道特性和噪声对基音检测的影响，从而提高基音检测的精度。通过理论模型验证了该方法的可行性，而且实际语音信号处理结果表明，该方法在基音检测时可以基本不受声道特性和噪声的影响。

关键词: 同态解卷, 基音检测, 语音, 频率不变线性滤波器