Abstract : A way to improve outputs produced by automatic speech recognition (ASR) systems is to integrate additional linguistic knowledge. Our research in this field focuses on two aspects: morpho-syntactic information and thematic adaptation. In the first part, we propose a new mode of integration of parts of speech in a post-processing stage of speech decoding. To do this, we tag N-best sentence hypothesis lists with a morphosyntactic tagger built to take into account the specificities of transcriptions. We reorder these lists by modifying the score computed by an ASR system at the sentence level to include morpho-syntactic information. Experiments done on French-speaking broadcast news (Ester corpus) exhibit a significant improvement of the word error rate. Besides, we establish the contribution of morpho-syntactic information to improve posterior based confidence measures. In the second more exploratory part, we are interested in thematically adapting the language model (LM) of an ASR system. We propose a scheme that enables us to specialize speech decoding in an unsupervised way. We first segment the studied document into thematically homogeneous sections. To this end, we develop a new probabilistic framework to integrate different modalities (lexical cohesion, acoustic clues, and linguistic markers) and show its relevance to improve segmentation. We then build adaptation corpora retrieved from the Web by using an innovative procedure. We finally modify the LM with these specic corpora and show that, on thematic sections that are manually selected, this method significantly improves the LM, even if the increase of the word error rate is slight.