基于段长分布的HMM语音识别模型.PDFVIP

1 Vol .32 No. 1 2004 1 ACTA ELECTRONICA SINICA Jan. 2004 HMM 王作英, 肖 熙 ( , 100084) : HMM , , , HMM ( DDBHMM) . , DDBHMM HMM ( 178% ) , DDBHMM , . : ; ; DDBHMM : TN91234 : A : ( 2004) Duration Distribution Based HMM Speech Recognition Models WANG Zuoyin , XIAO Xi (Dept. of Electronic Engineering, Tsinghua University , Beij ing 100084, China) Abstract: In order to overcome the defects of the duration modelin of homo eneous HMM in speech reco nition, a Duration Distribution Based HMM ( DDBHMM) is proposed in this paper based on a formalized definition of a lefttori ht inhomo eneous Markov model,which has been demonstrated that it can be identically defined by either the state duration or the state transition proba bilities. The speaker independent continuous speech reco nition experiments have shown that, by only modelin the state duration in DDBHMM , a si nificant improvement ( 178% error rate reduction ) has been achieved comparin with the classical HMM . The ideal properties of DDBHMM will ive promise to many aspects of speech modelin , such as the modelin of the state duration, speed varia tion, speech discontinuity and the inter frame correlation. Key words: duration; speech reco nition; DDBHMM , Poisson . 1 [ 10] . , HMM[ 1] , , ESHMM( Expanded State HMM) [ 11, 12] HMM Markov ,