Gaussian Model Based Multichannel Separation
Alexey Ozerov, Hirokazu Kameoka · 2018
This chapter presents audio source separation methods that are based on local Gaussian modeling of multichannel mixtures in the time-frequency domain. This family of source separation approaches allows integrating both spatial properties of the mixing system and spectral properties of the audio sources in a principled way within unified models. The parameters of such models can be efficiently estimated using various iterative algorithms and the source signals may be then reconstructed via multichannel Wiener filtering. Another attractive property of these approaches is that they are often suitable for both underdetermined and (over)determined mixing conditions and they generalize corresponding single-channel source separation approaches, thus bringing a continuum within the complex topology of audio source separation problems. We first present the main principles of Gaussian model based multichannel source separation followed by models overview, model estimation criteria and algorithms. We then present in detail some popular approaches.