多数据流隐马尔可夫模型的流权值优化方法  

Stream-weight optimization method of multi-stream HMMs

作　　者：秦伟[1] 韦岗[1] 

机构地区：[1]华南理工大学电子与信息学院,广州510640

出　　处：《计算机应用研究》2007年第11期100-102,共3页Application Research of Computers

基　　金：国家自然科学基金资助项目(60172048)

摘　　要：提出了一种流权值的优化方法。这种优化方法是基于似然比最大化准则和N-best算法。实验表明,这种新的方法即使在少量优化数据的条件下,也可以得到合适的流权值。而且,在不同的信噪比条件下,利用这种方法优化的多数据流隐马尔可夫模型,都可以有效、合理地融合音频和视频语音,提高语音识别系统的识别率。This paper proposed a novel stream-weight optimization method based on the likelihood-ration maximization criterion and the N-best algorithm. The proposed method had advantages that not only computational complexity was significantly reduced, but also audio-visual speech recognition performance was significantly improved by using a small optimization data set. Further experimental results demonstrate that the audio-visual speech recognition system provides significant enhancement of robustness in noisy environments.

关 键 词：双模语音识别 似然比最大化准则 流权值 

分 类 号：TN912.3[电子电信—通信与信息系统]