Etude comparative de modélisation de langage par bigrams et par multigrams pour la reconnaissance de parole
Yassine MAMI, Frédéric BIMBOT.

The use of stochastic ngram models has a long and successful history in the research community; nowadays ngrams are becoming quite common in commercial systems, as the market demands more robust and flexible solutions. This approach is particularly interesting for its effectiveness and its robustness, but limited to modeling only local linguistic structures. To overcome this limitation, we propose the use of models with variable length. In this paper we present the multigram language models and we integrate them in a speech recognition system. The experiments are carried out on a France Telecom's dialogue application for stock exchange.