改善线性预测系数倒谱抗噪声性能的方法

Approach to improvement of robust performance of linear prediction coefficient cepstral

线性预测系数倒谱(LPCC)是说话人辨认系统中较为有效的特征参数之一,但是该参数的抗噪性能不好,当语音中含有噪声时,系统的识别率明显下降。基于MATLAB软件,建立了一高斯混合模型(GMM)的说话人辨认系统,提出了特征参数加权窗口的方法。通过对多种加权窗口的正确识别率比较,发现对LPCC低阶参数的加窗提升,可以改善系统的噪声鲁棒性。MATLAB仿真结果显示,采用加窗后的系统识别率得到了明显改善。

The linear prediction coefficient cepstral (LPCC) is one of the effective feature coefficients, but the performance of LPCC is not good in noise environment. The right recognition rate of a speaker identification system goes down evidently when the speech con- tains noise. A speaker identification system was set up based on Gaussian Mixture Model (GMM) with MATLAB and a approach of the cepstal liftering window was used to improve robust of the feature coefficient. With compared different liftering windows in the result of right recognition rate, the low cepstal terms of LPCC make better robust performance in speaker recognition. The simulation result in MATLAB indicates the performance of the system with liftering window is further improved.