基于相对自相关序列MFCC特征的丢失数据带噪语音识别方法

Missing Data Speech Recognition in Noisy Environments Based on the Relative Autocorrelation Sequence MFCCs

提出了一种基于相对自相关序列(Relative Autocorrelation Sequences,RAS)MFCC(Mel-Frequency Ceps-tral Coefficient)特征的丢失数据带噪语音识别新方法。首先分析了环境噪声对RAS-MFCC的影响,提出了一种基于掩盖原理的不可靠分量检测方法;然后采用丢失数据(Missing data,MD)技术来消除畸变分量对识别过程的影响,实验结果表明,本文所提的识别方法可以在不同类型和信噪比的噪声环境中有效提高RAS-MFCC的识别率,并且其性能优于典型的基于滤波器组(Filter bank)语音特征的丢失数据语音识别方法。

A new scheme for noisy speech recognition by combining missing data technique and a robust speech feature based on RAS-MFCCs is presented. The influence of noises to the RAS-MFCCs is analyzed and a method for detecting the unreliable components is developed. Marginalisation approach is then used to eliminate the negative effect of the distorted components. Experimental results show that the proposed scheme can effectively improve the performance of the RAS-MFCCs under a wide range of SNR for different kinds of noises and be superior to the conventional missing data technique based on the filter bank features.