基于VQ/CDHMM的噪声环境下汉语口令识别研究  

Chinese Spoken Password Recognition in Noise Based on VQ /HMM

作　　者：黄玲[1] 潘孟贤[1] 

机构地区：[1]合肥工业大学计算机科学与信息工程学院,合肥230009

出　　处：《计算机工程与应用》2003年第28期106-108,161,共4页Computer Engineering and Applications

摘　　要：该文研究了基于改进VQ/HMM模型的语音识别方法,设计实现了基于该模型的汉语口令识别系统;研究了鲁棒性特征参数问题,提出了一些新的基于MFCC和LPCC的高维动态参数;分别进行了纯净语音和不同信噪比语音的识别实验,分析比较了不同类型特征参数、训练状态数和高斯混合度对该系统识别性能的影响。在此基础上得出了以下结论:在加性白噪声的情况下,使用高维动态参数明显提高了系统的鲁棒性;在汉语两字组的短语音(口令)识别中,状态数取4,混合度取3时实验结果较好;利用不同特征参数的优势,进行信息融合,是提高系统性能的一个很好选择。In this paper an effective Chinese order recognition system is constructed using improved VQ/HMM model.The robustness of feature parameters is also studied here and some new dynastic parameters with high dimension are presented based on MFCC and LPCC.Influence of parameter types,trained states and Gaussian mixture degrees on sys-tem performance is analyzed and compared on the basis of voice recognition experiment in clean and noisy environ-ment.The conclusions of this paper are shown as follows :the robustness of system is obviously improved by the means of dynastic parameters with high dimension in the additive white noisy environment ;performance of Chinese spoken password recognition system is superior when state number is four and Gaussian mixture degree is three;Information fu-sion using different parameters is an effective approach to improve the recognition performance of system.

关 键 词：语音识别 连续隐马尔可夫模型 特征参数 矢量量化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]