Masked positive semi-definite tensor interpolation

Abstract

Time-frequency constrained interpolation of audio has proven to be an effective technique in removing a wide variety of acoustic disturbances. Traditionally these techniques assume that the signal is stationary for the duration of the interpolation, which limits the types of disturbances that can be addressed. In this paper we propose masked positive semi-definite tensor factorisation followed by a novel form of multichannel spectral subtraction to solve the problem, and we demonstrate excellent results on some real-world examples. The proposed methods can remove disturbances that were previously considered highly challenging to interpolate, for example a burst of wind noise in a voice recording.