Audio source separation with multiple microphones on time-frequency representations

Abstract

This paper presents various source separation methods that utilize multiple microphones. We classify them into two classes. Methods that fall into the first class apply independent component analysis (ICA) or Gaussian mixture model (GMM) to frequency bin-wise observations, and then solve the permutation problem to reconstruct separated signals. The second type of method extends non-negative matrix factorization (NMF) to a multimicrophone situation, in which NMF bases are clustered according to their spatial properties. We have a unified understanding that all methods analyze a time-frequency representation with an additional microphone axis.