Robust speech recognition in adverse environments by separating speech and noise sources using JADE-ICA

Abstract

Spoken dialogue information retrieval applications are the future trend for mobile users. The typical presence of background noise in mobile environments causes significant reduction in speech recognition accuracy unless the recognizers are trained explicitly for each environment. However, it is not practical to train recognizers for every environment. One solution to this problem is to use an array of microphones and then separate the speech signals from noise by applying Blind Source Separation (BSS) techniques. BSS based on Independent Component Analysis (ICA) is one of the efficient techniques to separate mixed signals and JADE-ICA is one of ICA’s variants that is applicable for non-Gaussian signals. The JADE-ICA is applied here, since generally speech and noise are non-Gaussian signals. An array of two microphones was considered and two signal sources: speech and in-vehicle noise were assumed. Recognition accuracy experiments were conducted on 1831 utterances from 1831 speakers after separating the speech signals from noise using the JADE-ICA. The word recognition accuracy for three conditions: clean speech, noisy speech (SNR = 0 dB) and noise separated speech are 91.9%, 45.2% and 91.9%, respectively. This indicates that using the JADE-ICA to separate speech from noise results in the same recognition accuracy as that of clean speech. The details on JADE-ICA and simulations will be provided in the full paper.