PERFORMANCE IMPROVEMENT OF HIGHER-ORDER ICA USING LEARNING PERIOD DETECTION BASED ON CLOSED-FORM SECOND-ORDER ICA AND KURTOSIS

A hands-free system with conventional independent component analysis (ICA) should update the separation filter constantly to follow the hourly environment change. However, when the separationfilter is updated in the period where the user absent, ICA often yields an undesired separation filter. In this paper, we propose a novel algorithm of kurtosis-based voice activity detection (VAD) for an appropriate ICA optimization. In this algorithm, first, closed-form 2nd-order ICA (SO-ICA) is performed for providing the roughly separated signals, and based on the kurtosis of their signals, VAD is processed. Next, when the current time period is voiced block, the higher-order ICA’s re-optimization is applied using the previously obtained SO-ICA’s solution as the initial filter; this results in a fast and high convergence. The effectiveness of the proposed method is shown in a simulation experiment for blind spatial subtraction array with the proposed method.