[PAST EVENT] Improving Speech Classification in Home Environments Using Reverberant Environment Simulation

May 9, 2013
3pm - 4pm
McGlothlin-Street Hall, Room 2
251 Jamestown Rd
Williamsburg, VA 23185Map this location
Continuous care home monitoring systems improve healthcare management by providing information about activities of daily living. In management of psychological disorders like depression, speech is an important modality. However, current speech monitoring technologies for pertinent features require highly controlled monitoring situations. As a result, when such solutions are employed in more natural, ambient monitoring scenarios, they perform poorly. In this paper, we present our system RESonate and show how it can mitigate the effects of reverberation in real home or office environments for different speech classification applications e.g., speaker identification and mood detection. Our results show that reverberation has a deleterious effect on the performance of these applications. We have shown that our approach can improve performance considerably in the presence of reverberation using only very basic room dimension information and training data. RESonate minimizes a user's training efforts by requiring few samples of their voice to be recorded, which are subsequently transformed into a new reverbed training set using simple room acoustic models. This method of simulated reverb training was found to be accurate within 5-10% of the more time-consuming training performed in a real room. We evaluate RESonate using an emotional speech data set, collecting voice samples from four volunteers in four different rooms in homes and offices under controlled settings, and finally by deploying our system for successful six week long-term study in a home and office.

Bio: Robert F. Dickerson works in the areas of sensor networks, mobile computing, and acoustic processing. He is particularly interested in applications that address healthcare challenges. His smarthome system, Empath, featured in MIT Technology review, is currently being used in several clinical studies involving epilepsy and Alzheimer's disease, and soon for tracking depression. He is currently a Ph.D. candidate at the University of Virginia. He earned his B.S. in Computer Engineering at the University of Florida in 2006 and his Master of Computer Science degree in 2008.

Department of Computer Science