Subband gradient flow acoustic source separation for moderate reverberation environment

We present a subband source separation algorithm for miniature microphone arrays with dimensions smaller than the wavelength. By relating temporal and spatial gradients of the observed microphone signals in an anechoic environment, gradient flow converts the mixture of delayed sources to linear instantaneous mixture of the time-differentiated source signals, that can be then localized and separated using static linear independent component analysis algorithms. For source separation in multi-path environment, we propose subband decomposition of the spatial gradients estimated over an array of 4 microphones. The static ICA algorithms are applied in each frequency band and the localization results obtained from the ICA applied on the unfiltered spatial gradients resolve the scaling and permutation indeterminacy. The simulations with the room acoustic model and experimental results with conference room recordings demonstrate over 12dB separation in moderate reverberation environment.