基于单边自相关序列的语音特征及其在带噪语音识别中的应用

提出了一种抗噪声语音特征。首先计算语音信号单边自相关序列的差分序列，再计算该差分序列的线性预测系数，进一步求出例说系数。实验证明，传统的线性预测例谱系数和边自相关序列的一性预测倒谱数相比，采用单边自相关序列差分序列的线性预测倒谱系数作为语音信号的特征矢量，可以提高语音识别系统对带噪音语音的识别率。

Speech Feature Based on One-sided Autocorrelation Sequence and Its Application to Noisy Speech Recognition

The authors propose to use the linear predictve coding cepstrum coefficients (LPCC), which are extracted from the difference sequence of the one-sided autocorrelation sequence, as a noise-robust speech feature. Experimental results showed that the recogniser using the proposed feature can obtain higher recognition rate than that using the conventional LPCC or one-sided autocorrelation sequence LPCC in noisy environments.