Informations morpho-syntaxiques et adaptation thématique pour améliorer la reconnaissance de la parole

A way to improve outputs produced by automatic speech recognition (ASR) systems isto integrate additional linguistic knowledge. Our research in this eld focuses on two aspects:morpho-syntactic information and thematic adaptation.In the rst part, we propose a new mode of integration of parts of speech in a post-processingstage of speech decoding. To do this, we tag N-best sentence hypothesis lists with a morpho-syntactic tagger built to take into account the speci cities of transcriptions. We reorder theselists by modifying the score computed by an ASR system at the sentence level to includemorpho-syntactic information. Experiments done on French-speaking broadcast news (Estercorpus) exhibit a signi cant improvement of the word error rate. Besides, we establish thecontribution of morpho-syntactic information to improve posterior based con dence measures.In the second more exploratory part, we are interested in thematically adapting the lan-guage model (LM) of an ASR system. We propose a scheme that enables us to specialize speechdecoding in an unsupervised way. We rst segment the studied document into thematically ho-mogeneous sections. To this end, we develop a new probabilistic framework to integrate di erentmodalities (lexical cohesion, acoustic clues, and linguistic markers) and show its relevance toimprove segmentation. We then build adaptation corpora retrieved from the Web by using aninnovative procedure. We nally modify the LM with these speci c corpora and show that, onthematic sections that are manually selected, this method signi cantly improves the LM, evenif the increase of the word error rate is slight.