Spatio-temporal filtering methods for enhancement and separation of speech signals

摘要

In this paper, we give an overview of the background for, the ideas behind, and the challenges to be addressed in the project 'Spatio-Temporal Filtering Methods for Enhancement and Separation of Speech Signals,' which is funded by the Villum Foundation. The project aims at addressing the problem of enhancing and separating speech signals from noisy mixtures, a problem also known as the cocktail party problem. It aims at exploring new ways of solving this problem by generalizing a new class of optimal temporal filtering methods for periodic signals to multiple microphones, resulting in so-called spatio-temporal filtering methods that are controlled by two parameters, the direction-of-arrival and the fundamental frequency. These filters are optimal in that they let the signal of interest pass undistorted while everything else is attenuated as much as possible. Unlike state-of-the-art methods, they do not require knowledge of the statistics of noise and interfering speech signals, something that is especially important when dealing with non-stationary noise.