Design theory of microphone arrays for teleconferencing

Room reverberation and interfering acoustical noise lower the quality of speech transmission in teleconferencing. Conventional solutions for speech capture that suppress pickup of reverberation and interference typically constrain the motion of the speaker, encumber the speaker physically, or make it difficult for many different speakers to easily participate. These impediments result from cables, radio and headset microphones, or so-called house microphones located at fixed positions. Microphone arrays and matched-filter processing have been proposed as solutions to this sound capture problem. In part because the behavior of arrays in reverberant acoustic spaces has not been well quantified, there have been no guidelines for designing effective teleconference arrays. The principal goal of this work is the development of general design rules for speech acquisition arrays. Expressions for the array performance measures of signal-to-reverberation ratio and signal-to-interference ratio are derived using statistical acoustics, then verified by computer simulation using the image method of geometric room acoustics. The analysis is facilitated by assuming the reverberant sound field to be diffuse. The assumption is valid above approximately 250 Hz. All work assumes omnidirectional sources and sensors mounted on walls with low absorption. Array sensors should be placed close to, and roughly equidistant to the source, yet as far from each other as possible. The aperture of a full-band array should extend into three dimensions. Wall and ceiling mounting of sensors is recommended. Relations specify the number of sensors required for various room volumes, room absorptions, and source-to-sensor distances. A planar ceiling array expands the range of motion for speakers by widening the focal region above 300 Hz. Off-line audition of a 32-sensor array in a 60-cubicmeter room with a reverberation time of .63 second indicates that subjectively good array performance may be obtained using processing with matched filters truncated at 30 milliseconds. A listening test using audio from eight virtual arrays operating in a simulated room provides indications of their subjective performance relative to a single microphone operating at four different distances. An array with 32 wall-mounted sensors was constructed in an conference room. Off-line output of the array confirms the general principles of the design theory