Abstract
This paper describes the implementation of fast hidden Markov model (HMM) match algorithm in a phoneme-based Malay continuous speech recognition system. The decoding algorithm decouples the computations of state-likelihoods of phone HMM's from the main search which is bounded by syntactical and lexical constraints. This avoids the redundant state-likelihood computations of identical phone HMM's for different word models in the tightly integrated search and thus substantially reduce the decoding time. The algorithm is implemented in the framework of one pass dynamic programming search. For a 541-word task, the fast HMM match reduces the real-time factor (RTF) by a factor of 31.8 from 286.45 x RT to 9.02 x RT, compared to without decoupling. The word accuracy is maintained at 91.6% without loss for a test set perplexity of 15.45 in speaker-dependent mode.