Physics-based source separation using ambisonic microphones

We previously developed a physics-based method for separating two spatially distinct acoustic sources that overlap in time and frequency using acoustic vector sensors, devices that capture both acoustic pressure and particle velocity. Here, we extend the approach to ambisonic microphones, specialized microphones that capture 3-D spatial acoustic fields using closely spaced directional microphone capsules. The algorithm is generalized to operate on A-format/B-format signals, accounting for capsule orientation, spacing, and sensor mismatch. We demonstrate its ability to resolve overlapping arrivals, including correlated direct and multipath components of a tonal signal. Performance is evaluated through simulation and real loudspeaker measurements made in a reflective environment, with comparisons to conventional and adaptive beamforming methods such as delay-and-sum, MVDR, and MUSIC. This method offers a physically interpretable alternative to data-driven spatial filtering for complex acoustic scenes. Work sponsored by ONR TFO.