L'adaptation thématique d'un modèle de langue fait-elle apparaître des mots thématiques?

Gwénolé Lécorvé (1) , Guillaume Gravier (2) , Pascale Sébillot (1)

1 TEXMEX - Multimedia content-based indexing
2 METISS - Speech and sound data modeling and processing

Whereas topic-based adaptation of language models (LM) claims to increase the accuracy of topic-specific words within automatic speech recognition, this paper investigates why this wish is not always verified. After outlining the mechanisms of LM adaptation and automatic speech recognition, diagnosing elements are proposed along with solutions. In addition to a better accuracy on topic-specific words, results show better graph error rates and word error rates on a set of spoken documents with various topics

Domains: Signal and Image Processing; Signal and Image processing

Gwénolé Lécorvé, Guillaume Gravier, Pascale Sébillot. L'adaptation thématique d'un modèle de langue fait-elle apparaître des mots thématiques?. Journées d'Étude sur la Parole, May 2010, Mons, Belgium. ⟨inria-00555850⟩