语音信号的预滤波和倒谱参数方差的统计

对于非实验室条件下的实时语音识别系统来说，能否准确提取出所需要的语音信号特征，是决定该识别系统正识率高低的关键因素之一，对语音信号进行预滤波以抑制噪音，有利于语音信号特征的提取，本文结合一个实际系统，讨论了其降噪问题，并建议以倒谱参数方差之统计特性是否符合一定规律的为标准，来衡量预滤波的效果，文中还对数字滤波器及随机低门限时域中心削波在预滤波中的作用进行了讨论，实验结果表明，预滤波取得了较好的效果

On Prefiltering in Speech Recognition and Measuring of CCV

The problem of noise-reduction which arises in the design of a practical speech recognition system is discussed. Before any method is employed to reduce the noise, this paper proposes that the LPC cepstral coefficient variance (CCV) should be measured in addition to the measurement of RSN for judging whether the noise is negligible or not. Based on the judgement, a time-domain center clipper which follows successively a high-pass digital filter is employed for the noise-reduction task, although it is seldom used in noise-reduction before. The recognition test results are given to support the analysis in this paper. The error rates are about one-fourth of that obtained without the prefiltering processing.