Evaluation of time compression for connected word recognition

Recently, several studies have shown the interesting aspects of nonuniform sampling of the filtered speech signal in the context of an isolated word recognizer. This paper investigates the effect of nonuniform sampling for connected word recognition. Three nonlinear time compression techniques are evaluated, one that brings all reference utterances down to one same length, and others, for which the utterance length is variable. The nonuniform sampling approach is compared to the uniform one by opposing the three non-linear methods to two linear time compression ones. The results show that the variable length trace segmentation technique gives the best scores under all conditions, and that the uniform sampling approach can therefore be advantageously used in connected word recognition processes.