Computationally efficient audio segmentation through a multi-stage BIC approach

In this paper, we propose a computationally efficient approach for unsupervised audio stream segmentation via the Bayesian Information Criterion (BIC). Based on traditional BIC and DISTBIC, a novel multi-stage framework is presented. A statistic mean Euclidean distance based segmentation algorithm is used to pre-select candidate segmentation boundaries, and then delta-BIC integrating energy-based silence detection is employed to perform the segmentation decision to pick the final acoustic changes. Experimental results show that this method can greatly improve the whole detection process speed by a factor of 400 compared to that in Chen's while achieving a 19.2% reduction in the missed detection rate at the expense of a 3.8% increment in the false alarm rate using CCTV news data.