Combined-order hidden Markov models for reverberation-robust speech recognition

Abstract

In this contribution, the concept of combined-order hidden Markov models (CO-HMMs) is introduced by joining the first-order Markov and the second-order conditional independence assumption. The proposed approach is motivated and evaluated in the context of reverberation-robust automatic speech recognition. Two predecessor-dependent output probability density functions per hidden Markov model (HMM) state are employed in order to explicitly cope with the high inter-frame correlation in presence of reverberation. At the same time, the state duration modeling related to the first-order Markov assumption is addressed by a recently published training procedure based on hard alignment having the significant advantage that any conventional HMM can be efficiently updated to a CO-HMM. The experimental results show a reduction in average entropy as well as in word error rate in reverberant environments compared to conventional HMMs.