A Microphone-Array System for 
Speech Recognition Input 
Harvey F. Silverman 
Laboratory for Engineering Man/Machine Systems (LEMS) 
Brown University 
Objective 
This project is concexned with underlying mathematical algorithms, acoustics, hardware, and software to gain an 
understanding about, demonstrate tbe principles of, and, ultimately, to build an appropriate microphone-array system for 
speech-recognition input. 
Approach 
The approach taken might be called "recursive build-and-study". After investigating the layout problem and potential 
DSP architectures, we developed our first system. This allowed us to investigate real data and learn the real issues, begin to 
understand the difficult acoustics problems, and develop better DSP designs. This process is being repeated. 
Recent Accomplishments 
A new, nonlinear optimization algorithm called Stochastic Region Contraction (SRC), has been developed and has been 
applied to the microphone placement problem, talker location, and talker characterization. We have found that SRC is 
nearly two orders of magnitude faster than was simulated annealing. Our current research array system has been "har- 
dened", and recal-time, time-domain beamforming is operational. 
Plans for the coming year 
• Pefect one or more talker location algorithms and implement into real-time system 
• Develop adjunct algorithm to work with locator to track a talker 
• Demonstrate preliminary algorithms for talker characterization 
• Complete a second research platform which can perform both tracking and beam forming in real time 
• Develop accurate test methodology for understanding the acoustic properties of the sensor/room environment. 
410 
