基于频域和时域结合的基音周期提取算法

基音周期提取是语音编码和语音识别领域的一项重要研究课题。为了解决传统的自相关方法容易出现的半频倍频错误,提出了基于时域和频域分析的提取算法。该算法首先提取时域自相关值最大的若干个候选值;然后统计每个候选值对应的频域上所有相邻两个谐波能量和的最大值,用来对其自相关值进行加权;最后根据历史的基音周期值以及候选基音周期所对应的频域能量值对加权值进行修正。使用Keele数据库进行测试表明,使用该算法后基音周期提取的半频倍频错误率比传统算法下降了50%左右。

A pitch period detection algorithm using time and frequency analyses

Pitch period detection is an important part of speech encoding and speech recognition.Double and half pitch period detection errors in traditional auto-correlation methods are eliminated by combined time and frequency analysis.This algorithm first chooses some candidates based on the auto correlation function and calculates the maximum of all the near harmonics energy levels in the frequency domain for each candidate.Each energy maximum is modified based on its own energy level and the historical pitch periods as the weight of each auto correlation value.Simulations based on the Keele database show that the double and half error rates are reduced by about 50% relative to the traditional method.