基于MVDR和ICA的语音识别方法研究

Speech Recognition Based on MVDR and ICA

本文讨论了最小方差无失真响应建模方法,并与线性预测方法进行了比较,比较发现最小方差无失真响应滤波器能提供一个更好的原始语音包络。然后在研究ICA原理及FastICA快速算法的基础上,将MVDR参数提取方法与独立分量分析方法相结合,并与传统语音识别方法在有噪声和无噪声的情况下进行了比较,进而对识别率、计算时间等结果进行了分析。MVDR参数提取方法可以提高语音识别系统的识别率,但是会增加平均识别时间;而经过ICA特征变换后的语音识别系统具有较好的鲁棒性。

This paper discusses a modeling method of Minimum Variance Distortionless Response（MVDR）, which compares with the linear predictive method. The result is that the Minimum Variance Distortionless Response filter can afford better original speech envelope. On the basis of studying ICA and FastlCA, the MVDR parameter picking-up method is combined with independent component analysis, which compares with the traditional speech recognition under the conditions with noise and without noise, and the data are analyzed. The MVDR parameter picking-up method can improve the distin- guishing rate of the speech recognition system, but increases the mean recognition time; and the speech recognition system with the ICA characteristic transform is more robust.