K. Cho and Y. Yamashita (Japan)
Speech recognition, N-best, Recognition score, Heuristics
Recently statistical techniques have greatly improved the performance of speech recognition. The correct result is not always obtained when only the most probable recognition result is shown to the user. Therefore, some speech interface systems display several candidates of speech recognition to the user, and they ask him/her to choose the correct answer from the candidates. In such a speech interface system based on the N-best speech recognition, the determination of the number of candidates to be shown becomes an important problem. When many recognition candidates are displayed the probability that the correct answer is included will becomes high, but the user needs much time and much effort to search the correct answer. This paper describes a technique of determining the number of candidates dynamically using the distribution of the recognition scores for the N-best speech recognition. The proposed method reduces the number of candidates to be shown without degrading the speech recognition rate.
Important Links:
Go Back
