A method for modeling liaison in a speech recognition system for French.

In French the pronunciations of many words change dramatically depending on the word immediately preceding it. The result of this phenomenon, known as “liaison”, in an ASR system that does not model “liaison” is the requirement of unnatural pronunciation and much user dissatisfaction. We present, in this paper, the development of an acoustic model which takes into account the wide variability of word pronunciations caused by the liaison, the integration of this model into a French continuous speech recognition system and decoding results.