Top-down attention to complement independent component analysis for blind signal separation

Abstract

For robust speech recognition in real-world noisy environments, we present an algorithm to incorporate blind signal separation based on independent component analysis (ICA) and top-down attention processing. While ICA-based unmixing networks learn the inverse of mixing characteristics in frequency domain, their performance is limited by mismatches between the real-world mixing characteristics and assumptions of the ICA algorithm. The top-down process from a multiplayer Perceptron (MLP) classifier provides additional information on the speech signal, and fine-tunes the networks to compensate for the mismatches. For noisy speech signals recorded in a real office environment, the developed algorithm demonstrated great improvements on recognition performance.