Signal Processing for Sound Capture

High-quality electret microphones and single-chip processors are economical enough to be used in large numbers. This latitude opens opportunities for dynamic source location and sound capture with spatial selectivity in three dimensions, This report discusses algorithms for matched-filter processing of microphone arrays and for coordinate tracking of moving talkers. A prototype conferencing system is demonstrated in which the automatic source locator steers both a video camera and a beamforming microphone array to capture image and audio from a moving talker. In addition, a large scale array processing system is discussed. ~TRODUCTION Recent technology development in the field of telecommunication applications has placed increased demands on the acoustic capture capabilities of communication systems, The requirement of hands-free, untethered operation of telecommunication equipment often creates a situation where the user is at a significant distance from the microphone used for capturing her speech. The quality of the captured signal in such a configuration may be severely degraded by the effects of reverberation and additive noise. The microphone array provides a means of mitigating these effects. An array uses a set of spatially distributed sensors to provide signal-to-noise ratio (SNR) improvement for sound produced in the focal region. Earlier techniques such as beamforming [1] and approximations to inverse filters [2] treat reverberation as noise and seek to remove it from the captured signal. Matched Filter Array (MFA) processing [3] uses reverberant energy to reinforce the desired signal by cohering the reflected source images with the direct wavefront arrival. In order to properly focus a microphone array for sound capture, the location of the sound source must be known, Algorithms such as the cross-power spectrum phase (CPSP) [4,5] have been employed to estimate sound wavefront time-delay-of-arrival (TDOA) across selected microphone pairs. These estimates may be used to compute source location coordinates, This paper is organized as follows. First, the MFA algorithm for sound cap~re will be discussed, followed by algorithms for automatic source tracking. Then two platforms have been proposed that provide the required signal capture, signal processing, and device control capabilities for the implementation of practical microphone array system for teleconferencing use. Finally, conclusions and plans for future work will be presented. MFA PROCESSING FOR SOUND CAPTURE In an enclosure, sound propagation from source to microphone can be modeled as a room transfer function (RTF). The MFA algorithm consists of filtering the input signal obtained from each microphone with the time reverse of the focus-to-sensor RTF, For a sound source located at the focus, the effect of the matched filter is to convolve the undistorted signal with the autocorrelation of the focus-to-sensor RTF. The output of the MFA is simply the sum of the outputs of each individual matched filter. It is shown in [6] that the potential SNR of on-focus MFAprocessed speech is independent of the number of reflections present in the acoustical environment. In practice, however, this SNR does degrade with increasing reverberation time, but not as much as delay-sum beamforming [5]. In addition, due to computational constraints and perceptual issues, it is necessary to truncate the matched filters used for the processing. This algorithm was implemented offline using recordings made in real rooms. Using an 8-sensor array, an SNR improvement of about 6 dB over single-microphone recordings was seen in rooms with a 40-80 square meter floor area [5], with the sources between 2 and 8 meters away from the microphones. The focal region of the MFA was empirically found to be on the order of 10 cm. These results suggest that small-scale MFA processing can be effective in mitigating the effects of reverberation. SOURCE TRACK~G Source tracking is performed on windowed segments of captured microphone array signals. The CPSP method produces TDOA estimates across selected sensor pairs chosen from the array elements. The CPSP expression is a