A Voice Activity Detection Algorithm Using Sparse Non-negative Matrix Factorization-based Model Learning in Spectro-Temporal Domain

Voice activity detectors are presented to extract silence/speech segments of the speech signal eliminate different background noise signals. A novel voice detector is proposed in this paper using spectro-temporal features extracted from auditory model signal. After extracting scale, rate, and frequency feature space, a sparse structured principal component analysis algorithm used consider basic components these reduce dimension learning data. Then vectors employed learn models by non-negative matrix factorization algorithm. The procedure performed represent each vector with proper rate based on selected atoms. detection input frames computing energy representation for frame over composite model. If calculated exceeds specified threshold, it indicates that has structure similar atoms learned concludes observed content. results were compared other baseline methods classifiers processing field. These presence stationary, non-stationary periodic noises investigated they shown method can correctly detect activities.