Close Sound Source Localization incorporating Semi-Supervised Variational Bayesian NMF

This paper proposes a method to integrate a sound source separation method based on Nonnegative Matrix Factorization approach that is able to cope with partially known acoustic characteristics, and a sound source localization method with a frequency bin selection utilizing the separated signals. The proposed method is effective in the case when the target sound source locates close to a noise source, and the noise characteristics are available owing to the sound sound separation suppress the effect of noise. Numerical simulations showed that the proposed system succeeded to localize two sources of 5 degrees interval, which validates the approach.