Neural Voice Activity Detection and its practical use

Abstract

The task of producing a Voice Activity Detector (VAD) that is robust in the presence of non-stationary background noise has been an active area of research for several decades. Historically, many of the proposed VAD models have been highly heuristic in nature. More recently, however, statistical models, including Deep Neural Networks (DNNs) have been explored. In this thesis, I explore the use of a lightweight, deep, recurrent neural architecture for VAD. I also explore a variant that is fully end-to-end, learning features directly from raw waveform data. In obtaining data for these models, I introduce a data augmentation methodology that allows for the artificial generation of large amounts of noisy speech data from a clean speech source. I describe how these neural models, once trained, can be deployed in a live environment with a real-time audio stream. I find that while these models perform well in their closed-domain testing environment, the live deployment scenario presents challenges related to generalizability.