Cosine similarity based dictionary learning and source recovery for classification of diverse audio sources

A dictionary learning based audio source classification algorithm is proposed. Cosine similarity measure is used to select the atoms during dictionary learning. Three proposed objective measures, namely, signal to distortion ratio (SDR), the number of non-zero weights and the sum of weights have been used for classification. A frame-wise source classification accuracy of 98.86% is obtained for twelve different sources using SDR measure and a secondary support vector machine classifier. 100% accuracy has been obtained using moving SDR accumulated over 14 successive frames. For ten of the audio sources tested, 100% accuracy requires accumulation of only 6 frames of a signal.