Cluster-Based Discriminative Weight Training Framework for Voice Activity Detection

In this paper, a robust voice activity detection (VAD) for arbitrary noise environment is proposed. The conventional VAD has a limitation that the VAD is performed well in a particular environment. To cope with the limitation, speech and noise classes are divided into several clusters using unsupervised clustering, By discriminative weight training, optimal weights of each cluster are obtained and the weighted sum of the individual features is used for VAD. For performance evaluations, classification error rate is measured. The results show that the proposed method yields better performance than the conventional one.