摘要:
A speech recognition method can be preferably applied to equipment for constantly performing speech recognition, converts speech into an acoustic parameter series, calculates for the acoustic parameter series the likelihood of a hidden Markov model 22 corresponding to the speech unit label series about a registered word and the likelihood of a virtual model 23 corresponding to the speech unit label series for recognition of speech other than the registered word, and performs speech recognition based on the likelihoods.