Discriminative PLCA based polyphonic source identification

This work aims at searching for discriminatively learned features that characterize an audio source and make it identifiable even in polyphonic audio. Probabilistic latent component analysis (PLCA) is an effective method for decomposing a polyphonic signal into individual sources using source specific dictionaries. This work proposes a novel discriminative approach to find better PLCA dictionaries which discriminate between the pitched sources more efficiently, by learning the differences in their harmonic spectra. The experimental results for source identification show promising advantages of the proposed approach over the generative PLCA approach for source classification.