An algorithm is proposed on the basis of wavelet transform and waveform in time domain to improve the accuracy of pitch detection.The pitch period of clear speech and noisy speech in different signal-to-noise ratio is detected by using the proposed method and the traditional methods of autocorrelation function and average magnitude difference function.The experiments show that the half frequency errors may occur in the method of autocorrelation function and double frequency errors may occur in the method of average magnitude difference function.Furthermore, the error frames of the two methods increase with the decrease of signal-to-noise ratio.In the proposed algorithm, there are relatively few double and half frequency errors, and the pitch contour is clear and smooth without big jump.The results are coincident with the common rule as the slow variation with time of speech signal.Therefpre, the accuracy of pitch detection can be improved effectively by the algo rithm.%为了准确地检测语音信号的基音周期,采用小波变换和时域波形相结合的方法,分别用传统的自相关法、平均幅度差法及文中算法对纯净语音和不同信噪比时的含噪语音进行基音检测.实验表明,自相关法易出现半频错误,平均幅度差法易出现倍频错误,且两者随着信噪比的降低,错误帧数呈增加趋势.然而文中算法倍频、半频错误相对较少,基音轮廓清晰、平滑,无大的跳变,符合语音信号慢时变性的一般规律,从而提高了基音检测的精度.
展开▼
