Voice Activity Detection Based on Statistical Model Employing Deep Neural Network
Inyoung Hwang, Joon Hyuk Chang · 2014
In this paper, we propose statistical model-based voice activity detection (VAD) technique using deep belief network (DBN). From an investigation of the statistical model based VAD, it was discovered that the geometric mean of likelihood ratio as decision function is not desirable for the nonlinear input space and thus support vector machine (SVM) with nonlinear kernel function was proposed as the novel decision function. However, the SVM cannot be considered as strong one since it cannot fully take the nonlinear distribution of parameters, due to its shallow property. This problem can be addressed by the novel VAD framework using DBN which can fully fuse the advantages of multiple features through multiple-layer deep architecture. To achieve successful performance at statistical model-based VAD, we apply DBN as decision function. The performance of the proposed VAD algorithm is evaluated in terms of an objective measure and shows significant improvement compared to the conventional algorithms.