Using sparse representations for exemplar based continuous digit recognition

This paper introduces a novel approach to exemplar-based connected digit recognition. The approach is tested for different sizes of the exemplar collection (from 250 to 16,000), different length of the exemplars (from 1 to 50 time frames) and state-labeled versus word-labeled decoding. In addition, we compare the novel method for selecting exemplars, based on Sparse Classification, with a conventional K-Nearest-Neighbor approach. For word-labeled decoding we developed a Viterbi search that applies minimum and maximum duration constraints. It appears that Sparse Classification outperforms KNN, while state-labeled decoding provides better performance than word-labeled decoding. In all conditions the performance increases with the size of the collection. However, the optimal window length is 10 frames for state-labeled decoding, but 35 frames for word-labeled decoding.