说话人语音特征子空间分离及识别应用

Separation of speech feature subspace and its application to speaker identification

本文依据主元分析原理从语音特征观察空间分离说话人语音特征子空间,对输入语音特征矢量与子空间的距离测度进行了定义,并对基于特征子空间的说话人识别性能进行了分析。说话人语音训练样本提取特征后在语音特征观察空间形成具有一定散度的分布,根据主元分析原理和分布散度提取主要散度本征向量作为基底构成说话人语音特征子空间,并通过测试语音特征矢量与子空间的距离测度进行模式匹配。实验结果表明,特征子空间方法对说话人识别是有效的,特别是在小于3秒的短时测试语音下能够得到较高的识别率。

A new method for separation of speech feature subspace from observation space is proposed based on principal component analysis, and the performance of its application to speaker identification is evaluated. For every speaker, speech features extracted from training samples become a distribution with specific statistical properties such as mean and variance in observation space. Instead of statistical description, a feature subspace with the base of some significant eigen vector extracted from covariance matrix is constructed to describe speech feature distribution of speaker. Distance metrics for measuring distance between input feature vector and subspace are also proposed for pattern matching. Experiments on speaker identification performance analysis demonstrate effectiveness of subspace method, especially for short time test speech with length less than or equal to 3 seconds.