基于一种新特征参数的说话人识别

Speaker recognition based on a new feature parameter

针对传统的特征参数Mel频域倒谱系数MFCC难以满足语音信号的非平稳性问题,提出一种基于小波分析的新特征参数FPBW的提取方法.为了提高训练速度,采用正交高斯混和模型,将正交变换改到最大期望EM算法之前进行,从而减少训练时间.实验结果表明,新的特征参数FPBW优于特征参数MFCC,并且采用正交高斯混合模型进一步提高了识别性能和训练速度.

Aimed at the problem that the traditional feature parameters MFCC （reel-frequency ceptrum coefficients） was hard to satisfy the non-stationary characteristic of speech signal, a method was proposed for extraction of a new feature parameter FPBW based on wavelet analysis. In order to improve training speed, an orthogonal Guass mixture model （OGMM） was employed in order that the orthogonal transform was to be performed before the use of expectation maximization algorithm, so that the training time was reduced. The experiment results showed that a new feature vector FPBW was better than MFCC, and the OGMM could further improve the recognition performance and training speed.