Advanced Voice Activity Detection on Mobile Phones by Using Microphone Array and Phoneme-Specific Gaussian Mixture Models

This paper presents an advanced voice activity detection (VAD) system, developed for mobile Android OS platforms with limited hardware capabilities. The system uses a dual microphone array for noise suppression and a decoder with a constrained grammar for speech detection, where Gaussian mixture models (GMMs) are used together with their acoustic weights and energy in order to increase the robustness of the proposed system. The system is presented as part of the Voice Assistant application for mobile phones, and the results are given on a database that was especially designed for that purpose. The results presented in this paper show a high accuracy even when a large amount of background noise is present.