7th International Conference on Spoken Language Processing
September 16-20, 2002
In this paper, we examine the idea of new information integration in the acoustic models of false starts and Out-Of-Vocabulary (OOV) words with the aim of reducing the global error rate. The method consists in reproducing partially the acoustic form of false starts and OOV words in the related rejection model. We present for the moment the integration of information about the phonetic length. We estimate from the training corpus the phonetic length distributions of these events and we use the associated probabilities to constraint the output of the rejection model. The obtained results were compared to those of a system based on a standard phonetic loop model. We observed a 4.7% absolute reduction in the global Word Error Rate (WER).
Bibliographic reference. Hamimed, H. / Damnati, G. (2002): "Integration of phonetic length properties in the acoustic models of false starts and out-of-vocabulary words", In ICSLP-2002, 865-868.