Princeton's neighbor-consistent filters slash sound zone instability by 61.8%
Audio zones stay stable even with jittery head tracking — 55.9% less filter variation.
Personal sound zones (PSZ) let multiple listeners share a space without headphones, but they rely on accurate head tracking. Small tracking errors — from optical distortion, occlusions, or sensor jitter — cause loudspeaker filters to jump wildly, ruining the audio experience. Now, Hao Jiang and Edgar Choueiri from Princeton's 3D Audio and Applied Acoustics Lab have developed a solution: neighbor-consistent neural filters.
The team trains coordinate-conditioned neural networks to generate real-time PSZ filters, but adds a regularization term that penalizes large filter differences between a given listener position and its randomly perturbed neighbors. This forces the network to produce smooth, stable outputs even when tracking coordinates are noisy. In simulations using a 25-position anchor set and a split-band system, the method slashed RMS spatial variation rates by 55.9% in the woofer band and 30.3% in the tweeter band. Real-world tests with a 24-driver linear array and two stationary head-and-torso simulators showed worst-case neighborhood isolation improved by 16.9%, and spatial variation rates dropped up to 61.8% — all while maintaining strong isolation between zones. The work makes PSZ systems dramatically more resilient to the inevitable imperfections of real-world tracking hardware.
- Neighbor-consistency regularization reduces RMS filter variation by up to 55.9% (woofer band) and 30.3% (tweeter band) in simulation.
- In situ with a 24-driver array, worst-case isolation improves 16.9% and spatial variation rates drop 61.8%.
- Method preserves isolation quality while making personal sound zones robust to common tracking noise like jitter and occlusion.
Why It Matters
Makes head-tracked spatial audio practical for noisy real-world environments, enabling stable multi-listener experiences.