基于多相滤波器组的语音基频检测方法

基于多相滤波器组的语音基频检测方法，运用多相滤波器组分解语音信号频谱，然后利用声带震动的能量准周期性在各子带进行峰值搜索，并综合这些子带的搜索结果计算基音周期，最后根据先验知识以及一种新的清浊音判定方法对结果进行校正。基于标准TIMIT语料库子集的实验结果证明了此方法的有效性。

A More Accurate Pitch Detection Algorithm (PDA)

Past PDAs, to our best knowledge, were unable to predict accurately the pitch at any instant. We now propose a PDA based on multi phase filter bank that can do so. Section 1 discusses in much detail the design of multi phase filter bank. Essentially it explains how to decompose the speech spectrum for removing vocal tract effects. Fig.1 shows the prototype low pass filter. Section 1 also gives the procedure of bandwidth selection. Section 2 discusses in much detail the procedure for pitch detection. Essentially it performs peak searching in each sub band and gets decision for each sub band; it synthesizes all the sub band decisions to obtain the final detection results. Section 2 also gives the rules for correct peak searching. Section 3 proposes a new method for making voiced/unvoiced decision based essentially on the following fact: the energy spectrum of the voiced speech is different from that of the unvoiced speech. Experimental results given in section 4 on sentences chosen from TIMIT database show preliminarily that out new PDA based on multi phase filter bank can predict accurately the pitch at any instant.