Acoustic scene characterization in wireless sensor networks.

This dissertation provides an in-depth study on the application of wireless sensor networks to speaker-related acoustic environments. A set of system requirements are established including hardware configuration for the network nodes, requirements for real-time operation of acoustic sensing, and the capabilities for distributed information processing and wireless data routing. A distributed sensing and classification system is developed, analyzed and implemented for acoustic scene characterization. In particular, the characterization of voice-related environments is focused on the presence of multiple speakers. Voice signals of interest acquired from multiple sensors are utilized to determine the number of speakers, their gender, and their emotional state. This application is of interest in several security applications. In terms of signal analysis techniques, algorithms are formulated and implemented for single and multiple acoustic target detection. Also, methods that exploit acoustic correlations between voice and speaker monitoring are implemented and decision fusion is employed to improve pertinent estimates. Commercial off the shelf hardware platforms are utilized to form the real-time testbed for experiments. The testbed includes algorithm integration and the real-time implementation. For the gender classification problem, a Bayesian framework is exploited to develop sensor and decision fusion algorithms. Locally observed information is combined to achieve a more reliable global decision. The proposed fusion techniques are robust to additive background noise at different levels. Simulation and real time results show that important speaker information is estimated and refined through the use of distributed sensing.