基于独立感知理论的鲁棒语音识别算法  

Robust speech recognition algorithm based on fletcher-allen principle

作　　者：孙暐[1] 吴镇扬[1] 

机构地区：[1]东南大学无线电工程系,南京210096

出　　处：《东南大学学报（自然科学版）》2005年第4期506-509,共4页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金资助项目(60272044);国家重点基础研究发展计划(973计划)资助项目(2002CB312102).

摘　　要：为了提高在噪声环境下语音识别系统的性能,对基于子带独立感知理论的语音识别方法进行了研究.这些方法利用人耳对不同频率信号感知的差异,以及噪声和识别对象的频域特征差异,分别采用线性分析、判决分析、多层感知机以及子带最大似然估计对噪声影响进行补偿.实验表明,子带分析采用非线性策略优于线性策略.基于独立感知假定的子带模型,虽然由于独立性假定丢失了带间相关性,但对于噪声环境下语音识别而言可以捕获噪声和识别对象的频谱差异,从而获得比全带分析更高的鲁棒性.To improve the robust of.the speech recognition systems, the subband robust speech recognition algorithms based on the Fletcher-Allen principle are studied. These algorithms utilize the perception difference of human＇s ear to the different frequency signal and the spectrum difference between noise and recognition objects to compensate the effects of noise with maximum likelihood criteria and linear combination or discriminative combination or multi-layer perceptron. The test shows nonlinear analysis is superior to linear analysis. Although the subband model based on the Flether-Allen principle loses the correlation among the different bands, it can catch the difference between noise＇s spectrum and the recognized object＇s spectrum to achieve higher robust than the whole band model in noisy environments.

关 键 词：语音识别 隐马尔可夫模型 最大似然 

分 类 号：TN912.34[电子电信—通信与信息系统]