The Joint LIMSI and Vecsys Research Systems for NBEST 2008
Julien Despres, Petr Fousek, Jean‐Luc Gauvain, Yvan Josse, Lori F Lamel, Abdel Messaoudi · 2008
This document describes the speech recognizers jointly submit-ted by the LIMSI and Vecsys Research to the N-Best 2008 eval-uation. The aim of this evaluation was to perform automatic speech recognition (ASR) for the Dutch language. Northern Dutch and Southern Dutch (also known as Dutch- NL, and Flemish- VL) have been processed with two different data types per accent (broadcast news- BN, and conversational telephone speech- CTS). The speech recognizers use multiple decoding passes with models (lexicon, acoustic models, language mod-els) trained for the four different transcription tasks: BN-NL, BN-VL, CTS-NL and CTS-VL. Four primary systems (one for each accent-domain task) have been trained for the primary training condition and the unlimited decoding. The primary submission is also a less than 10xRT contrastive submission. Four contrastive systems have a processing time of 1xRT. The case-sensitive word error rates (WER) of the primary LIMSI-Vecsys Research systems on the N-Best development data are