基于复杂性的说话人识别技术探讨

Studies on Speaker Recognition Based on Complexity

目的用非线性方法中的复杂性特征来分析非平稳的语音信号。方法采用 5 1位被试者讲同一段话的语音数据 ,语音样本包括两类 :同一被试者在不同时刻讲同一段话及不同被试者讲同一段话 ,分别从复杂性特征曲线和复杂性特征曲线重构的嵌入空间分析处理。结果同一被试者不同时刻讲同一段话的复杂性曲线间的差异 ,明显地低于不同被试者讲同一段话的复杂性曲线间的差异。另外 ,从复杂性特征曲线重构的嵌入空间可以看出 ,同一说话人的复杂性曲线离散点相对集中于在此嵌入空间中同一区域 ,而与其他人的特征离散点存在统计学上的明显差异。结论复杂性分析方法能够用于语音特征分析 。

Objective To analyze the speech characters with computation complexity. Method Voices of 51 testees spoke the same paragraph were recorded and the same sentence of voice waveform was intercepted as source data. There were two kinds of sample voices: same testee speaking the same sentence at different time and different testee speaking the same sentence. Result The computation complexity curves of the different testee were obviously distinguishing, while those of the same testee were almost the same. In a 2D embedded space the computation complexity features of individual testee differs with others even if testees speak the same sentences. Conclusion This complexity features might be applied to speaker recognition system and using complexity method to analyze speech signals has wide application prospect.