基于离散小波变换和感知频域滤波的语音特征参数

A New Speech Feature Extracted by Wavelet Analysis & Mel-Frequancy Filtering

为了提高语音识别系统对环境噪声的鲁棒性，本文通过对几种特征参数的噪声鲁棒性的研究,在感知倒谱分析（Mel-Cepstrum）的基础上，引入了多分辨率小波分析技术,利用小波变换的时频局域性并结合感知频域上的滤波技术，提出了一种新型的语音特征参数（DWT-MFC），该特征参数对于环境噪声具有很强的耐噪性。实验表明，与（LPCC）和（MFCC）参数相比，采用DWT-MFC参数使系统的噪声鲁棒性得到显著的提高。

In order to improve the noise robustness of speech recognition systems , the robustness of several kinds of speech features have been studied extensively. Based on Mel-Cepstrum Analysis and the Multi-scale Wavelet Analysis, a new kind of speech feature (DWT-MFC) is discovered. This kind of feature has great immunity to the environmental noise. The experiment results show that the system built on DWT-MFC has better noise robustness than on LPCC and MFCC.