广州话语音声纹鉴定中的量化分析研究

Quantitative analysis in speaker identification

本研究从一般元音(音节核心)的同一认定方法着手,着重从定量分析方面研究声学特征参量,把多元统计方法使用到声纹鉴定中。通过计算变异系数、相关系数、正态性检验、方差分析等基本统计分析,对于男性语音,制定了马氏距离判别数学模型,判对率达98.3%;对于女性语音,制定了费歇(Fisher)判别数学模型,判对率达93.5%。本研究解决了声纹鉴定对声学特征参量定量分析方面的问题,确定了判定阈值及鉴定所需音节数,提供了一种新的语音判别方法。

In this study, a new method of vowel （vowel nuclear） identification by multivariate statistical analysis, indued analysis on coefficient of variation, coefficient of correlation, test for normality, and analysis of variances （ANOV） was introduced, we construct a identification model of Mahalanobis distance method for male speakers with the correct rate of 98.3 %, and a model of Fisher＇ s discriminant method for female with the correct rate of 93.5 %.