This paper describes the utilization of a forward heuristic function to achieve fast and accurate beam search in HMM-LR continuous speech recognition. We also discuss "dynamic rejection" as a means to reject out-of-grammar utterances during the search in the same framework. The key idea is utilizing the forward heuristic function to compensate for the time differences among the partial hypotheses. It has been experimentally shown that this approach drastically reduces computational cost. Moreover, it offers parallel processing capability with speech input to realize a real-time system. The new concept in dynamic rejection is the "0th hypothesis", whose the accumulated likelihood function is defined as the forward heuristic function. By using the "0th hypothesis" score as the base score, dynamic rejection operates by pruning in the framework of beam search. The ideas presented in this paper are widely applicable to time-asynchronous approaches to speech recognition with reduced computational cost.
Bibliographic reference. Noda, Yoshiaki / Sagayama, Shigeki (1995): "Fast and accurate beam search using forward heuristic functions in HMM-LR speech recognition", In EUROSPEECH-1995, 913-916.