Acoustic Source Direction by Hemisphere Sampling Stanley T

Acoustic Source Direction by Hemisphere Sampling Stanley T. Birchfield Daniel K. Gillmor Quindi Corporation Palo Alto, California

The Problem sound source q compact microphone array f 5/11/01

Previous Work two microphones f [Huang et al. , ICASSP 1999; Stephenne and Champagne, ICASSP 1995] four microphones q, f [Brandstein et al. , ICASSP 1995; Brandstein and Silverman, CSL 1997] large microphone arrays r, q, f [Brandstein et al. , ICASSP 1995; Svaizer et al. , ICASSP 1997] 5/11/01

Our microphone array setup microphone d=15 cm Note: Arbitrary compact configurations can be handled 5/11/01

Principle of Least Commitment: “Delay decisions as long as possible” Example: 5/11/01

Previous Algorithm: Cone Intersection mic 1 signal decision is made early prefilter correlate mic 2 signal find peak prefilter intersect mic 3 signal (may be no intersection) prefilter correlate mic 4 signal q, f find peak prefilter [Brandstein et al. , ICASSP 1995; Brandstein and Silverman, CSL 1997] 5/11/01

Our Algorithm: Sampled Hemisphere mic 1 signal prefilter correlate mic 2 signal prefilter correlate map to common coordinate system … sampled hemisphere final sampled hemisphere combine q, f find peak correlate mic 3 signal prefilter correlate mic 4 signal prefilter map to common coordinate system temporal smoothing decision is made after combining all the available evidence 5/11/01

Pair-wise matching of signals microphone mic 1 and mic 2 mic 3 and mic 4 all four mics 5/11/01

Sampled Hemisphere asymptotical cones (black regions are theoretical “blind spots”, where cones have no intersection; practical blind spots are larger) sampled hemisphere 5/11/01

Maximum error from matching non-coincident microphones additional robustness outweighs possible error, which is negligible (less than 5 degrees when r > 4 d) 5/11/01

Experimental results (two trials, A and B) A B A B Note: also has been extensively used (hundreds of hours) in real environments 5/11/01

Summary Advantages of sampled hemisphere algorithm – handles arbitrary microphone array configurations – has no “blind spots” – demonstrates increased robustness, by following the principle of least commitment Future work – integrate into a sound source location system – investigate multiple simultaneous sound sources 5/11/01