一种语音特征提取中Mel倒谱系数的后处理算法

摘要

为提高语音识别系统的鲁棒性，本文以Mel频率倒谱系数（ MFCC）为基础，结合均值消减法、方差归一化、时间序列滤波法和加权自回归移动平均滤波法，提出了一种后处理算法，本文将该算法命名为MVDA后处理法，所得语音特征参数简称MVDA。本文首先从理论上推导了MVDA后处理法可以去除加性噪声和卷积噪声的干扰，接着针对MVDA与MFCC做了对比试验，并分析了含噪语音与语音信号的欧氏距离变化，证明MVDA后处理法的每一步均有效降低了噪声的干扰，且得出了MVDA在不同噪声环境中均更优的结论。这种简洁的语音特征不仅可以达到许多复杂语音特征处理方法的效果，而且有效减少了自动语音识别系统的计算量。%To improve the robustness of automatic speech recognition systems, a new speech feature postprocessing method based on the Mel⁃frequency Cepstral Coefficient ( MFCC) is proposed, which is named the MVDA postpro⁃cessing method. The postprocessed feature parameters are named MVDAs. This technique combines mean subtrac⁃tion, variance normalization, time sequence fltering, and autoregressive moving average flters. Experiments were conducted to compare MVDA and MFCC. Changes in the Euclidean distance of the speech with noise and the speech signal were analyzed, proving that every step of MVDA postprocessing could effectively reduce the noise in⁃terference. Thus, all MVDAs in different noise environments were superior. This simple feature does not only a⁃chieve the effect of many complex speech feature processing methods but also effectively reduces the computational complexity of automatic speech recognition systems.