基于多带HMM和神经网络融合的语音识别方法的信道鲁棒性  

Channel Robust of Speech Recognition Based on Multi-band HMM and BPNN Fusion

作　　者：姚志强[1] 戴蓓倩[1] 李辉[1] 黄伟[1] 

机构地区：[1]中国科学技术大学电子科学与技术系,合肥230026

出　　处：《计算机工程与应用》2004年第1期71-73,82,共4页Computer Engineering and Applications

基　　金：安徽省自然科学基金项目资助(编号:01042205)

摘　　要：对于基于HMM的语音识别系统,由于训练和测试环境(背景噪声。语音传输信道Microphone等)的失配将会造成识别性能的严重下降。根据人类的听觉感知机理,该文针对语音传输信道失配问题,提出了一种基于多带HMM的系统结构,有若干个子带系统和一个全频带子系统组成,并采用神经网络对个子系统的输出进行后端融合及判决。实验表明,这种方法可以有效地提高识别系统的信道鲁棒性。Due to the mismatch between trai ni ng and testing environment (such as background noise,speech transition channel-Microphone),automatic speech recognition(ASR)based on HMM tends to drastic ally degrade in performance.For the reason of the theory of human aural percept ion,in this paper,we present a system based on multi-band HMM,composed of several sub-band systems and a full-band system.Then,we use ANN to make the final decision by combined all systems' outputs.Through test,our system i s proved to be efficient to upgrade the performane of channel robustness of the recognition system.

关 键 词：信道失配 子带HMM 神经网络融合 信道鲁棒性 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]