Probabilité a posteriori: amélioration d'une mesure de confiance en reconnaissance de la parole
Julie Mauclair, Yannick Estève, Paul Deléglise.

This paper adresses the word posterior probability used as a confidence measure on speech recognition system. We present a new confidence measure based on the behavior of language model back-off used during the recognition processing. Merging this new confidence measure with word posterior probability allows to obtain a fusion confidence measure, called WP/LMBB, which outperfoms the word posterior probability. Our experiments have been carried out on the corpus used during ESTER, the french evaluation campaign on automatic transcription of french broadcast news. Using the normalized cross entropy (NCE) as an evaluation metric, which is used in NIST evaluations, experimental results on test data of ESTER evaluation show a very significant improvement: whereas the word posterior probability reaches a value of NCE equal to 0.187, the WP/LMBB measure obtains 0.270.