Exemplar selection techniques for sparse representations of speech using multiple dictionaries

This paper describes and analyzes several exemplar selection techniques to reduce the number of exemplars that are used in a recently proposed sparse representations-based speech recognition system. Exemplars are labeled acoustic realizations of different durations which are extracted from the training data. For practical reasons, they are organized in multiple undercomplete dictionaries, each containing exemplars of a certain speech unit. Using these dictionaries, the input speech segments are modeled as a sparse linear combination of exemplars. The improved recognition accuracy with respect to a system using fixed-length exemplars in a single dictionary comes with a heavy computational burden. Due to this fact, we investigate the performance of various exemplar selection techniques that reduce the number of exemplars according to different criteria and discuss the links between the salience of the exemplars and the data geometry. The pruned dictionaries using only 30% of the exemplars have been shown to achieve comparable recognition accuracies to what can be obtained with the complete dictionaries.