A fast compressive sensing approach for phoneme classification

In this paper, a new fast compressive sensing (CS) algorithm for phoneme classification is introduced. In this approach, unlike common CS classification approaches that use CS as a classifier, we use CS as an N-best class selector to limit the secondary classifier input into certain classes. In addition, we use a tree search strategy to select most similar training set for the specific test sample. This makes the system adapted to each test utterance and reduces the empirical risk. By this approach, we obtain promising results comparing with other well known classifiers. In addition, the employed CS approach is a fast l⁰ norm algorithm which dramatically reduced the computational complexity in the recognition phase.