Voice Activity Detection based on Combination of Multiple Features using Linear/Kernel Discriminant Analyses

This paper presents a voice activity detection (VAD) scheme that uses multiple of some popular features. As, in each noisy condition, one type of feature performs best in speech/non- speech classification, combining features can lead to a better performance. Features are combined linearly with weights that were obtained for each condition in training stage via a method of classification and dimension reduction, i.e. linear discriminant analysis (LDA). Also, nonlinear combination of features is carried out via kernel discriminant analysis (KDA). The results are compared to MCE-based approach. We show that LDA and KDA can lead to better performance, compared to MCE and also the best single feature. In particular, KDA shows 12.14% improvement in average speech/non-speech classification rate, relative to the best feature.