语音识别中的加性噪声补偿研究  

Research on speech recognition compensation for additive noise

作　　者：田斌[1] 田红心[1] 易克初[1] 

机构地区：[1]西安电子科技大学综合业务网国家重点实验室,陕西西安710071

出　　处：《西安电子科技大学学报》2001年第3期292-295,共4页Journal of Xidian University

基　　金：国家自然科学基金资助项目!( 69872 0 2 7)

摘　　要：针对语音识别中的加性噪声进行研究 ,提出了动态自适应多模板谱减法和多模板谱加训练补偿法 .动态自适应多模板谱减法是多模板谱减法和噪声动态自适应方法的有效结合 ,使谱减法既能适应环境中存在多种噪声的情况 ,又能有效利用当前得到的噪声环境信息 .而多模板谱加补偿法从谱减法的逆向角度 ,并基于多模式训练的思想 ,有效地使模板适应多种噪声情况 ,该方法不增加识别时的运算量 ,有利于考虑噪声的各种复杂情况 .This paper proposes two methods for speech recognition under the additive noise environment, namely dynamic adaptation multi-model spectral subtraction and multi-model spectral addition. Taking advantages of both multi-model spectral subtraction and noise adaptation, the dynamic adaptation multi-model spectral subtraction method can not only cope with the environment of several types of noise, but also make use of the dynamic information of the background. The multi-model spectral addition method is derived from the opposite of the conventional spectral subtraction, which can accommodate the model with all sorts of noise. Moreover, the compensation is made during the training process and does not increase the complexity of the recognition procedure.

关 键 词：语音识别 加性噪声 谱减法 语音信号处理 

分 类 号：TN912.34[电子电信—通信与信息系统]