Abstract:Voice activity detection (VAD) plays an important role on the performance of speech processing systems. Recently, more and more works focused on the statistical model-based VAD algorithms have been presented in literature, which make a decision of speech and nonspeech based on the likelihood ratio (LR). However, all the statistical models used in those algorithms are unable to exactly describe the statistics of noisy speech and various type noises. In this paper, a novel VAD algorithm is proposed based on the nonparametric detection theory by incorporating the likelihood ratio into the sign test to provide a new decision rule. Meanwhile, an optimal threshold of the proposed method is derived and the selections of relevant parameters are discussed as well. Experimental results show that the proposed VAD algorithm outperforms the conventional statistical model-based VAD.

Robust Statistical Voice Activity Detection Using a Likelihood Ratio Sign Test.