Instrument identification using PLCA over stretched manifolds

Probabilistic latent component analysis (PLCA) is a popular tool for decomposing the spectra of polyphonic music for identifying the constituting musical instruments. Supervised PLCA reconstructs the observed polyphonic spectra using instrument specific spectral parts. However, the distance metrics between two instrument classes in the spectral space may not be the same as that in the acoustic space. This paper proposes a novel geometrical approach to PLCA by modifying the metrics in a way so as to enhance the inter-class distances, while reducing the intra-class distances. Hence, the likelihood function of PLCA is made to give more weight to reconstructing the features which discriminate the sounds of different instruments better. Experiments on binary instrument classification give encouraging results with the proposed approach.