USING SPARSE REPRESENTATIONS FOR EXEMPLAR BASED CONTINUOUS DIGIT RECOGNITION
Gemmeke, et al.
We have extended our previous work on isolated digit recognition by applying Sparse Classification to continuous digit recognition. The non-parametric technique is based on the idea that arbitrary speech signals can be represented as a linear combination of suitably selected exemplars. The classification is based on finding the smallest number of labeled exemplars in a very large library of exemplars that jointly approximate the observed speech token. We applied a sliding time window approach by applying SC to every window individually and decoding the utterance using Viterbi decoding. We show that our method outperforms KNN.