Abstract: The wavelet transform and Wiener filtering are combined in order to de-noise and Modified Maximum Classification Error（MMCE） is used for speaker recognition.Noisy signals are decomposed into different subband by Daubechies wavelet of three sizes.Then these signals are de-noised by Wiener filter in every subband.The output of every subband is resumed by wavelet reconstruction.Finally wavelet coefficient is transformed into MFCC.A kind of modified MCE model to decrease computer multiplications and further to enhance computer velocity is proposed.Experiment result shows that the proposed method can decrease the computation multiplications and improve the identification rate of the system.

Key words: wavelet transform, Wiener filtering, Modified Maximum Classification Error（MMCE）

摘要： 将小波变换和维纳滤波结合起来对语音进行去噪和MMCE对说话人进行识别。说话人识别近来的关注点主要集中在子带处理的使用上。通过三尺度的Daubechies小波把输入含噪信号分解于不同子带中，然后在各个子带分别通过维纳滤波去噪，再把各个子带的输出通过小波重构恢复信号，最后通过Mel滤波器组把小波系数转换成MFCC（美尔倒谱系数）。提出了一种改进的MCE模型去减少计算量，并进而提高运算速度。实验结果显示：提出的方法减少了计算量，而且提高了系统的辨认率。

关键词: 小波变换, 维纳滤波, 改进的最大分类错误