Abstract

Nonnegative Matrix Factorization (NMF) is a powerful tool for decomposing mixtures of non-stationary signals in the Time-Frequency (TF) domain. In the literature, a variety of probabilistic models involving latent variables have been designed for introducing some a priori knowledge (like harmonicity and smoothness) into NMF. However, phases are generally ignored in such models, which results in a limited spectral resolution (sinusoids in the same frequency band cannot be properly separated). Moreover, most of these models assume that all TF coefficients are independent, which is not the case of sinusoidal signals for instance. In this talk, I will present a unified probabilistic model called HR-NMF, which achieves a high spectral resolution by taking both phases and local correlations in each frequency band into account. The potential of this new approach will be illustrated in the context of audio source separation and audio inpainting.