Discriminative sparse representation for speech mode classification

In this paper, we propose a discriminative sparse feature for speech mode classification (SMC). We use sparse representation, to calculate efficient representations of speech signal while learning dictionary atoms for speech mode classification. In this study, we collectively consider the three broad speech modes: conversation, extempore and read speech for mode classification. The usefulness of sparse representation depends on the dictionary. Therefore, we presented a Fisher discriminative (FD) based dictionary to obtain the variation in the speech mode. The feature extraction scheme using Fisher discriminative dictionary learning (FDDL) method is shown by exploring the discriminative information present in both coding coefficients and reconstruction error. The proposed sparse feature is extensively evaluated on Bengali speech corpus, and it results in the better performance than the conventional speech features.