Mesures de confiance trame-synchrone
Résumé
This paper presents some confidence measures for large vocabulary speech recognition that can be evaluated directly within the first steps of the recognition process. Having some clues to drive the recognition process may help to improve the accuracy of a sentence. Confidence measures may fit to this goal, so we propose some measures that can help the recognition as early as possible, without having to wait for the recognition process to be completed. Furthermore, our confidence measures are local, and they are based on partial word graphs. Experiments on a French broadcast news corpus are presented, and give results close to the post calculated version of the measures.