摘要
传统的隐马尔可夫模型的缺点在于不能很好地描述语音信号的动态特性．某些改进算法引入状态持续时间进行修正，但是也削弱了对实时信号长度变化的适应性．作者在传统的隐马尔夫模型的基础上，通过在引入状态持续时间时，将其归一化，并观察序列长度对它的影响，使之能较好地描述语音信号的动态特性，同时也能较好地自适应描述实时语音信号的长度变化．讨论了具体的算法，并给出了实验数据．结果表明，采用归一化状态持续时间隐马尔可夫模型的语音识别系统能较好地改善传统的隐马尔可夫模型在动态特性方面的不足．
Traditional hidden Markov model has long been blamed for its deficiency to express dynamic characteristics of speech signals.Some revisions try to improve the model by using state duration to modity the transition probability parameters.But they also result in poor adaptability to length variation of the real time signals.In this paper,authors introduce state duration.This leads to an improvement of its adaptability as well as its dynamic characteristics.In the rest of this paper,authors present the related algorithm and experimental results.It has been shown that the model is more effective to modelling in speech recognition system.
出处
《四川大学学报（自然科学版）》
CAS
CSCD
北大核心
1999年第5期857-863,共7页
Journal of Sichuan University(Natural Science Edition)
关键词
隐马尔可夫模型
状态持续时间
语音识别
hidden Markov model
adaptability
normalized state duration
speech recognition system
