Most existing rejection techniques for connected speech recognition are based on the likelihood score of different recognition hypotheses. The rejection procedure described in this paper attempts to improve the ability to reject out-of-vocabulary utterances combing likelihood measures with application-dependent knowledge. This knowledge is integrated in the recognition system using the N-best paradigm. The paper describes the adaptation of an N-best algorithm to our recognition task of telephone numbers. Two simple utterance rejection techniques are presented: one considering only the N-best recognition results and other introducing external knowledge. Experimental results using these techniques show that simple methods for utterance verification provide moderate performance using garbage models and the results from the N-best hypotheses. When external knowledge sources are included for our recognition task we obtain a 20% increase of performance in terms of recognition error with an excellent trade-off between false rejection and false alarms.
Bibliographic reference. Caminero-Gil, F. Javier / Torre-Munilla, Celinda de la / Hernandez-Gomez, Lúis / Martin del Álamo, Cesar (1995): "New n-best based rejection techniques for improving a real-time telephonic connected word recognition system", In EUROSPEECH-1995, 2099-2102.