Conclusion

In this paper, we described an effective speech recognition algorithm that uses word trigram models and a filled-pause procedure to reduce the memory and computational requirements. With these methods, we can perform execution in a 15M byte space for about a 1500-word vocabulary. Using this algorithm, a sentence recognition rate of 66.7% was obtained for speaker dependent recognition. We also described a procedure to skip $/pause/$ in speech data. Using this procedure, a sentence recognition rate of 83.9% was obtained as opposed to 0.0% with no such procedure for speaker independent recognition. This procedure was then extended to the procedure for dealing with filled-pauses in spontaneous speech. In experiments for spontaneous speech, we obtained a 42.0% sentence recognition rate; including the semantically correct sentences, the sentence recognition rate was about 75%. In the future, we will try to raise the sentence recognition rate for spontaneous speech by controlling the duration of filled-pauses.

Acknowledgments

We would like to thank Dr. Yamazaki, President, ATR Interpreting Telecommunications Research Laboratories and Dr. Sagisaka for their continuous support of this work. We are also grateful to all of the members of Department 1 for their advice and encouragement.