基于定量递归分析的清浊音判决

Voiced/unvoiced Decision Based on Recurrence Quantification Analysis

在语音信号处理中,清浊音判决的准确与否直接关系到后续语音处理的质量。该文通过分析不同的语音音素动力学物理模型在其递归图上的表现,统计定量递归分析中确定性和归一化最长对角线这两种特征参数,得到清浊音的显著差异。设定灵活合理的阈值判决语音信号的清浊音,得到了良好的试验结果。和其他传统判决方法比较,误判率有明显降低,为语音特征提取和识别研究提供了新的途径。

Voiced/unvoiced decision is an important component in speech signal processing. In this paper, different topological structures in Recurrence Plots （RPs） are described for the different physical models of speech production. By statistically analyzing the determinism and the normalized maximal length of diagonal structures acquired from Recurrence Quantification Analysis （RQA）, a flexible and efficient decision framework is proposed. Comparing with some traditional methods, the proposed algorithm has lower wrong decision rate. The method provides a new way for feature extraction and speech recognition.