SPEECH RECOGNITION METHOD AND SPEECH RECOGNITION APPARATUS

A speech recognition method in which, upon speech recognition with use of a model composed of subwords such as triphones depending on a plurality of context, inhibiting hypotheses from extending according to the plurality of contexts so that a processing rate of the speech recognition is increased. In computing a likelihood of an acoustic model referred to upon computation of an accumulated likelihood of a hypothesis through the speech recognition, with respect to each hypothesis, at least one subword model determined depending on its context subwords is referred to, likelihoods of a speech to be recognized and acoustic models of the respective subword models are computed, and the maximum value of the computed likelihoods is set as a likelihood of an acoustic model corresponding to the hypothesis.