THE LIMSI2006TC-STAREPPSTRANSCRIPTIONSYSTEMS*

Thispaperdescribes thespeech recognizers developed to transcribe European Parliament Plenary Sessions (EPPS) inEnglish andSpanish inthe2ndTC-STAREvaluation Campaign. Thespeech recognizers arestate-of-the-art systems using multiple decoding passes withmodels (lexicon, acoustic models, language models) trained forthedifferent transcription tasks. Compared totheLIMSITC-STAR2005EPPSsystems, relative worderror rate reductions ofabout 30%havebeenachieved onthe2006development data. Theworderror rates withthe LIMSIsystems onthe2006EPPSevaluation dataare8.2% forEnglish and7.8%forSpanish. Experiments withcross-site adaptation andsystem combination arealso described. Index Terms - Speech recognition