Experimental study on applying adaptation techniques for deep neural network based voice activity detection

Abstract

In this paper, we employ several model-spaceadaptation techniques for deep neural network (DNN)based voice activity detection (VAD) to adapt themodel to unseen background noise conditions.Adaptation results are evaluated in terms of theobjective measures such as frame accuracy, speechhit rate (HIT), and false alarm rate (FA).Experimental results suggest that the adaptation hasbeen carried out mainly to learn the noise-specificcharacteristics, rather than modeling the speechrelated features of the unseen adaptation utterances.