Multichannel Clustering and Classification Approaches
Michael Mandel, Shoko Araki, Tomohiro Nakatani · 2018
Time-frequency masks provide a powerful framework for source separation. In the multichannel setting, these masks can be constructed using the spatial characteristics of the sources and the observations in addition to their time-frequency characteristics. This chapter describes clustering and classification methods for doing so. Clustering methods group together time-frequency bins with similar characteristics. Clustering methods can be further divided into narrowband and wideband approaches, depending on whether the separation is coordinated across frequency. Classification methods predict source labels for time-frequency bins based on exposure to related training data. The chapter also discusses the estimation of the number of sources involved in a mixture, as well as methods for performing spatial filtering based on masks.