A real time speaker verification demonstration on the smart flow system

摘要

This paper summarizes our research and development effort on the investigation of an innovative approach to making a smart meeting room accessible to hands-free users with special needs. More specifically, we developed a real time system which is able to acquire speech from a privileged speaker at a distance across a noisy room, using a microphone array, and identify the speech signals from the privileged speaker, among others, in real time. We were successful in achieving this goal, and established that we can use commercially available speaker dependent speech recognition products to determine what the privileged speaker said, while simultaneously rejecting speech from other meeting attendees. To accomplish this goal, we implemented a real time speaker verification algorithm, based on Gaussian mixture models (GMM) commonly employed in speaker identification research. We used the smart flow system, developed by the National Institute of Standards and Technology (NIST), to conduct our experiments.