JiN'ichi Murakami and Shoichi Matsunaga
This paper describes an effective recognition algorithm that uses word trigram models directly and a procedure for dealing with filled-pauses in spontaneous speech. This recognition algorithm greatly reduces the memory requirements and computational costs by employing two techniques: beam search and an improved Viterbi search. With these methods, execution can be performed in a 15M byte space for about a 1500-word vocabulary. The filled-pause procedure, capable of handing many filled-pauses in spontaneous speech, is then examined for spontaneous speech recognition. Even though the proposed algorithm employs a simple procedure, a 42.0% sentence recognition rate is obtained for spontaneous speech. Including the semantically correct sentences, the sentence recognition rate is about 75%.