基于短时能量的循环AMDF基音检测算法  

Circular AMDF Pitch Detection Algorithm Based on Short-Time Energy

作　　者：马莎莎[1] 戴曙光[1] 穆平安[1] 

机构地区：[1]上海理工大学光电与计算机工程学院,上海200093

出　　处：《计算机仿真》2014年第7期278-282,共5页Computer Simulation

摘　　要：在语音识别优化问题的研究中,基音周期是语音信号的重要特征参数,准确提取十分困难。传统自相关函数(ACF)、平均幅度差函数(AMDF)等基音检测算法在基音周期提取中容易出现估计结果减半或者加倍的情况,导致检测结果准确率降低。针对上述问题,提出了短时平均能量和循环平均幅度差函数(CAMDF)相结合的基音周期检测方法。实验结果表明,改进方法在有效分离清浊音的基础上,加强了浊音基音周期所处采样点的谷值特性,减少了估计结果的半倍频错误,提高了基音周期估计精度,性能优于传统算法。Pitch period is a key characteristic parameter of speech signal, but accurate extraction of the period is very difficult. It usually occurs that halving and doubling frequency errors appear in the estimated results during the pitch tracking of traditional algorithm like the auto correlation function （ACF） and the average magnitude difference function （AMDF）. To solve this problem, a detection method of combining the short-time average energy with the circular average magnitude difference function （CAMDF） for pitch period is proposed. Experimental results indicate that, on the basis of effectively separating the voiced speech from the unvoiced speech, this method can increase the valley features of sampling point which indicates the voiced speech pitch period, reduce the halving, and doubling er- rors and improve the precision of estimated values. The performance of this method is better than that of the traditional algorithms.

关 键 词：基音周期检测 短时能量 平均幅度差函数 循环平均幅度差函数 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]