Duration-Distribution-Based HMM for Speech Recognition

Abstract

To overcome the defects of the duration modeling in the homogeneous Hidden Markov Model (HMM) for speech recognition, a duration-distribution-based HMM (DDBHMM) is proposed in this paper based on a formalized definition of a left-to-right inhomogeneous Markov model. It has been demonstrated that it can be identically defined by either the state duration or the state transition probability. The speaker-independent continuous speech recognition experiments show that by only modeling the state duration in DDBHMM, a significant improvement (17.8% error rate reduction) can be achieved compared with the classical HMM. The ideal properties of DDBHMM give promise to many aspects of speech modeling, such as the modeling of the state duration, speed variation, speech discontinuity, and interframe correlation.