5th International Conference on Spoken Language Processing
For large vocabulary continuous speech recognition based on hidden Markov models, we often face the issue of trade-off between the accuracy and the speed. A new method is proposed in this article such that complex models are used to retain a high accuracy whereas the speed is achieved by using the similarities in acoustic matches. These similarities are based on the assumption that we refer as a look-phone-context property. By using the look-phone-context property, the number of acoustic matches can be substantially reduced in the course of scoring all possible phonetic transcriptions of recognition hypotheses. Experiments on the speaker-independent Wall Street Journal task show that a fast-response system can be reached without compromising the accuracy.
Bibliographic reference. Lee, Clark Z. / O'Shaughnessy, Douglas (1998): "A new method to achieve fast acoustic matching for speech recognition", In ICSLP-1998, paper 0208.