ICA’s suitability assisted by Voice Activity Detection

Abstract

This research presents an innovative system for adaptive speech denoising using Independent Component Analysis (ICA) and Voice Activity Detection (VAD). Designed for instantaneous mixtures (two sources and two microphones), the proposed system identifies the noise contained in each noisy mixture. For that type of noise applies the most suitable ICA method among three methods (FastICA, Kernel ICA and JADE) and, after source separation, identifies the estimated speech signal. The signal mixtures are non-linear and the proposed system extracts information that can be used for further pre and/or post-processing. The experimental data shows that adaptive ICA allows better performance than applying a fixed ICA method for all hypothetic cases (an average ofldB SNR improvement). The process is completely automatic from the source recording to its output and such system has a wide range of applications.