Conference Paper
Physically Informed Acoustic Modeling for Neural Personal Sound Zone Rendering
AES Show 2026 Nashville · Paper 53 · October 2026
Abstract
Deep learning-based Personal Sound Zones (PSZs) commonly rely on simulated acoustic transfer functions (ATFs)
for training, yet idealized point-source models often exhibit a large sim-to-real gap. Recent work has shown that
adding physically informed components can improve sim-to-real generalization, but the individual contribution
of each component has remained unclear. This paper presents a controlled cumulative ablation study on a headpose-
conditioned binaural PSZ renderer based on the recently proposed Binaural Spatial Audio Neural Network
(BSANN). We progressively enrich simulated ATFs with three components: (i) anechoically measured frequency
responses of the particular loudspeakers (FR), (ii) analytic circular-piston directivity (DIR), and (iii) rigid-sphere
head-related transfer function (RS-HRTF). Four training configurations are evaluated with in-situ measurements
using two dummy heads at a single static pose. Performance is assessed using frequency-dependent inter-zone
isolation (IZI), inter-program interference (IPI), and crosstalk cancellation (XTC), summarized as broadband
log-mean values over 10020,000 Hz. Results show that FR primarily serves as a practical spectral calibration,
yielding a modest but consistent XTC improvement for both listeners (+0.87/+0.44 dB) and reducing inter-listener
IPI imbalance. DIR provides the most consistent gains in sound-zone separation (IZI +1.02/+0.59 dB; IPI
+0.79/+0.81 dB), achieving the best average IZI/IPI (10.05 dB). RS-HRTF inclusion dominates binaural separation,
boosting XTC by +2.38/+2.89 dB and increasing average XTC from 4.51 to 7.91 dB, primarily above 2 kHz,
while introducing mild listener-dependent shifts in IZI/IPI. These findings offer practical guidance on prioritizing
measurements and models when constructing training ATFs under limited budgets.
