基于Bark子波变换和RAS-MFCC特征的语音识别系统研究

Speech Recognition System Based on the Bark Wavelet and RAS-MFCC Features

提出了一种适用于低信噪比下的语音识别算法。该算法在前端利用人耳听觉掩蔽效应进行语音增强,然后提取一种在不需要噪声先验知识,能避免大量计算的情况下能较好地消除系统加性噪声的自相关MFCC(RAS-MFCC)参数,作为语音特征参数进行识别。

A new algorithm for a speech recognition system is presented to meet the demand for quality noise reduction algorithms capable of operating at the low signal-to-noise ratio. This algorithm combines the speech enhancement based on spectral subtraction which is improved by Bark wavelet transform, and the new robust features RAS-MFCC which can suppress the additive and convolution noise effectively without the need for prior knowledge.