关联小波神经网络与高斯混合模型说话人识别

Speaker Recognition Based on Gaussian Mixture Model Associated with Wavelet Neural Network

针对GMM模型在进行说话人识别时对噪声敏感以及在分类方面存在的缺陷,提出了一种小波神经网络和GMM结合的说话人识别模型,把GMM输出的似然概率和小波神经网络的训练相关联,采用带动量的BP算法和EM算法对小波神经网络和GMM模型分别训练,使目标说话人模型似然概率达到最大,进而提高说话人识别的效果。实验结果表明,新模型兼具小波神经网络抗噪声性能、学习分类能力以及GMM对说话人特征的描述能力,在多种噪声背景下能有效的提高说话人识别效果。

The Gaussian Mixture Model （GMM） applied in the task of speaker recognition is very sensitive to noise and has some defects in the aspects of classification.To solve these problems,a new speaker recognition model combines Wavelet Neural Network （WNN） and GMM was put forward in this study.The new model has both of the WNN＇s ability of anti-noise,learning classification and GMM＇s ability of describing speakers charac teristics.To improve the speaker recognition performance,the new model integrated the GMM＇s output likelihood probability with the WNN training.By adopting the momentum BP algorithm and EM algorithm to train WNN and GMM respectively,the likelihood probability in the new model was maximized.Experiments showed that the new model presented in this paper could effectively improve the speaker recognition performance in a variety of noise backgrounds.