Modèles de langage ad hoc pour la reconnaissance automatique de la parole. (Ad-hoc language models for automatic speech recognition)

The three pillars of an automatic speech recognition system are the lexicon, the language model and the acoustic model. The lexicon provides all the words that can be transcribed, associated with their pronunciation. The acoustic model provides an indication of how the phone units are pronounced, and the language model brings the knowledge of how words are linked. In modern automatic speech recognition systems, the acoustic and language models are statistical. Their estimation requires large volumes of data selected, standardized and annotated. At present, the Web is by far the largest textual corpus available for English and French languages. The data it holds can potentially be used to build the vocabulary and the estimation and adaptation of language model. The work presented here is to propose new approaches to take advantage of this resource in the context of language modeling. The document is organized into two parts. The first deals with the use of the Web data to dynamically update the lexicon of the automatic speech recognition system. The proposed approach consists on increasing dynamically and locally the lexicon only when unknown words appear in the speech. New words are extracted from the Web through the formulation of queries submitted toWeb search engines. The phonetization of the words is obtained by an automatic grapheme-to-phoneme transcriber. The second part of the document presents a new way of handling the information contained on the Web by relying on possibility theory concepts. A Web-based possibilistic language model is proposed. It provides an estition of the possibility of a word sequence from knowledge of the existence of its sub-sequences on the Web. A probabilistic Web-based language model is also proposed. It relies on Web document counts to estimate n-gram probabilities. Several approaches for combining these models with classical models are proposed. The results show that combining probabilistic and possibilistic models gives better results than classical probabilistic models alone. In addition, the models estimated from Web data perform better than those estimated on corpus.