Improvement of the Initialization of ICA Time-Frequency Algorithms for Speech Separation

摘要

The blind separation of speech signals in reverberant environments is a well-known problem for which many algorithms have been developed. In this paper, we propose a novel initialization procedure for those ICA algorithms that work in the time-frequency domain and use the prewhitening of the observations. In comparison with classical initializations, this method allows to reduce drastically the number of permutations. The effectiveness of the proposed technique in realistic scenarios is illustrated by means of simulations.