汉语连续语音识别中多项式拟合语音轨迹模型的研究  

Research on Polynomial-Fitting Speech-Trajectory Model in Chinese Continuous Speech Recognition

作　　者：欧智坚[1] 王作英[1] 

机构地区：[1]清华大学电子工程系,北京100084

出　　处：《电子学报》2003年第4期608-611,共4页Acta Electronica Sinica

基　　金：8 63基金 (No 2 0 0 1AA1 1 4 0 71 )

摘　　要：尽管作为当前最为流行的语音识别模型 ,HMM由于采用状态输出独立同分布假设 ,忽略了对语音轨迹动态特性的描述 .本文基于一个更为灵活的语音描述统计框架—广义DDBHMM ,提出了一个具体的多项式拟合语音轨迹模型 ,以及新的训练和识别算法 ,更好地刻划了真实的语音特性 .本文还给出了一种有效的剪枝算法 ,得到一个实用化模型 .汉语大词汇量非特定人连续语音识别的实验表明 。Although as the most popular model for speech recognition, HMM takes no account of the dynamics of the speech trajectory, since it assumes the outputs of a state to be independent and identically distributed. In this paper, based on a more flexible statistical framework for speech description-the generalized DDBHMM, a particular polynomial-fitting speech-trajectory model is proposed with new algorithms for training and recognition. It describes the real characteristics of speech more reasonably. With the effective path-pruning algorithm additionally proposed, it becomes a practicable model. Experiments on Chinese large-vocabulary speaker-independent continuous speech recognition show that with this path-pruned polynomial-fitting speech-trajectory model, the recognition performance is improved distinctively at relatively low computational cost.

关 键 词：连续语音识别 隐马尔可夫模型 基于段长分布的隐马尔可夫模型 

分 类 号：TN912.34[电子电信—通信与信息系统]