基于动态单边自相关序列和频率规整线性预测的抗噪声语音识别

Robust speech recognition based on dynamic one-sided autocorrelation sequence and frequency warped linear predictive coding

提出了一种既符合人耳听觉特性又具有良好抗噪性的语音特征分析方法。首先将单边自相关函数序列进行时间方向的平滑处理,提高单边自相关函数的抗噪性,然后用平滑后的单边自相关函数序列代替原信号进行频率规整的LPC分析,最后经倒谱变换得到该特征参数。数字语音识别实验证明:利用该特征参数的语音识别系统的识别性能优于MEL倒谱系数、LPC倒谱系数等传统的语音特征参数。

A representation of speech that invariant to noise is introduced. The idea is to filter the temporal trajectories of short time One-Sided Autocorrelation Sequence (OSAS) of speech such that the noise effect is removed. The filtered sequences are denoted as Dynamic Autocorrelation Sequences (DAS). Then frequency warped LPC (WLPC) algorithm is applied to the DAS instead of the original speech. This speech feature set, which not only corresponds to the performance of human auditory property, but also improves the noise robustness of speech recognition, is denoted as DAS-WLPCC. Chinese digit recognition experiment based on continuous density HMM shows the effectiveness of DAS-WLPCC features in presence of white noise and color noise.