Discriminative sparse-based feature extraction and dictionary learning for sound classification applications

This paper presents a novel sparse-based classification algorithm for audio applications such as sound classification. We propose performing the sparse feature extraction, the dictionary learning, and classification processes simultaneously. This discriminative learning procedure for adapting the dictionaries and classifier to each specified audio task, instead of employing pre-defined dictionaries is the main novelty of our work. According to our experiments, applying this algorithm on some Mel-scale spectral features, such as MFCC (Mel Frequency Cepstral Coefficient), instead of raw temporal data can improve the accuracy and execution time significantly. Our proposed discriminative MFCC-sparse features when evaluated on real data consisting of five audio classes, substantially out-performed the non-discriminative ones. The lengths of test segments in our method are less than 0.5 second. This potential of usage for real-time applications is another advantage of our proposed approach.