A Statistical Model-Based Voice Activity Detection Using Multiple DNNs and Noise Awareness

In this paper, we propose the ensemble of deep neural networks (DNNs) by using acoustic environment classification for statistical model-based voice activity detection (VAD). Since conventional decision functions for statistical model-based VAD are based on shallow model and it cannot take an advantage of the diversity of the space distribution of features, we present to use the multiple DNNs separately trained on different noise condition as decision function for the statistical model-based VAD. And, environmental noise classification is also performed based on the separate DNN since acoustic environment classification makes it possible to achieve high detection performance at various type of noise environment by using different algorithm according to current noise condition. In the training stage, a number of DNNs are independently trained according to different type of noise environments, and separate DNN is organized to detect one of the environmental conditions. In an online stage, the environmental knowledge on each frame is contributed to allow us to combine the speech presence probabilities, which are derived from the ensemble of the trained DNNs for the individual environment. Our approach for VAD was evaluated in terms of objective measures and showed significant improvement compared to the conventional algorithm.