Studies have shown that the segmental discriminative power of Neural Networks (NN) can be very useful to improve the performances of temporal sequence decoders like Hidden Markov Models (HMM) [1,2]. In this paper, we present a new connectionist segmental approach applied to the reordering of the N-best solutions provided by a HMM. The global system uses a segmental recognition framework, where phonetic segments are provided by the alignement of a speech utterance on a HMM. The scores of each solution are obtained by a two-level architecture. The first one is a «One Net One Class» connectionist architecture which provides phonetic scores for each phoneme belonging to a word, where phonetic scores can be interpreted as measures of validity of each segment labelling. The second level computes each word score as a product of the phonetic scores. In a final step, the NN scores and the HMM scores are combined for each of the N-best solutions in an optimal way in order to minimize word classification errors. The present system, with the knowledge of the 5-best solutions, leads to a 15 to 20% reduction of the error rate when compared to the one obtained with the HMM alone on several speaker-independent databases recorded over telephone networks.
Keywords: N-best solutions, segmental connectionist process- ing, segment labelling validation