This paper investigates the modelling of speech signal stationarity changes in order to improve both recognitions and rejection performances of an HMM based speech recognition system on field data, which also implies the rejection of noises and out-of-vocabulary utterances. Implemented in an N-best solution post-processing, the use of discrete probability distributions of the number of speech signal stationarity changes inside each phonetic segment enables the computation of a segmental postprocessing score. The experiments conducted on afield recording database resulted in a 6 % reduction in the substitution error rate on the correct and truncated utterances, and a 35 % reduction in the false alarm error rate on noises and out-of-vocabulary utterances. This yielded a 20 % global error rate reduction.