A MRF-based parallel processing algorithm for speech recognition using linear predictive HMM

Abstract

Parallel processing in speech recognition is described, which is carried out at each frame on time axis. The authors have already proposed a parallel processing algorithm for HMM-based speech recognition using Markov random fields (MRF). The parallel processing is realized by modeling the hidden state sequence by an MRF and using the iterated conditional modes (ICM) algorithm to estimate the optimal state sequence given an observation sequence. However this parallel processing algorithm is applicable only to the standard HMM but not to the linear predictive HMM which takes into account the correlations between nearby observation vectors. The authors propose a parallel processing algorithm applicable to the correlation-considered HMM, where a new deterministic relaxation algorithm is used instead of the ICM algorithm for estimation of the optimal state sequence. >