基于幅度差平方和函数的基音周期提取算法

为了在任意采样率下都可以高效、准确地进行基音周期提取,提出基于归一化幅度差平方和函数的基音周期提取算法。导出高效计算幅度差平方和函数的方法,时间复杂度是O(N lbN),给出该函数的归一化定义。归一化幅度差平方和函数的取值反映语音信号的非周期性程度,由此定义了基音周期的状态损失函数和转移损失函数,从而能在后处理过程中利用V iterb i算法,确定最优的基音周期序列。实验结果表明:与通用基音提取算法相比,在保证实时性的基础上错误率降低了9.31%,证明使用该算法提高了基音周期提取的准确率。

Real-time pitch tracking based on sum of magnitude difference square function

A pitch tracking algorithm was developed based on the normalized sum of the magnitude difference square function(SMDSF) for accurately estimating speech pitch at any sample rate in real time.The SMDSF can be calculated efficiently by FFT with a time complexity of O(N ln N).A normalized form of the SMDSF is related to the ratio of the aperiodic power to the total power.Thus,the state loss function and the transition loss function based on the normalized SMDSF can use the Viterbi algorithm to find the optimal pitch path.Test results show that the pitch tracing algorithm works in real-time with 9.31% less pitch estimation errors compared with the baseline pitch tracking system,which illustrates the accuracy of the normalized SMDSF-based pitch tracking combined with the Viterbi algorithm.