This paper presents a new statistical segmental post-processing of the N-best solutions obtained by a hidden Markov model based speech recognizer. This post-processing computes, for each solution, a new segmental score which is latter combined with the markovian score. The solution with the highest score identifies the answer. We introduce a new method in the statistical modelling of each segment involving two models: one for correct segmentation and one for incorrect segmentation. The first results presented here shows that this new approach leads to a 10 to 15 % reduction of the error rate.
Keywords: N-best solutions, segmental post-processing, statistical modelling.