EVALITA 2009: Abla Automatic Speech Recognizers Description

In this paper we describe the two systems we presented at the EVALITA 2009 workshop, for the connected digits recognition task. The former is an Abla srl proprietary speech recognizer, based on standard decoding algorithms, with syllabic acoustic models. The recognition phase is followed by a rescoring session, based on syllables energy and duration templates, which recover some recognition errors. The latter recognizer is a system, based on standard algorithms and triphonic acoustic models. The core comes from Nuance ASR 8.5 recognizer, trained on a proprietary corpus. Performances are discussed and compared to the best performing EVALITA 2009 system on this task.