Decomposing the Influence of Physical Acoustic Modeling on Neural Personal Sound Zone Rendering: An Ablation Study

Researchers Hao Jiang and Edgar Choueiri have published a detailed ablation study titled 'Decomposing the Influence of Physical Acoustic Modeling on Neural Personal Sound Zone Rendering.' The work, posted on arXiv, systematically breaks down how incorporating real-world physics into the simulated training data for deep learning-based Personal Sound Zone (PSZ) systems affects performance. The team used a head-pose-conditioned binaural renderer based on the Binaural Spatial Audio Neural Network (BSANN) architecture and evaluated four different configurations through in-situ measurements with two dummy heads. The core problem addressed is the 'sim-to-real' gap, where AI models trained on idealized point-source acoustic transfer functions (ATFs) fail to generalize to physical loudspeakers and human listeners.

The study progressively enriched simulated ATFs with three key components: the anechoically measured frequency responses of the actual loudspeakers (FR), an analytic circular-piston directivity model (DIR), and rigid-sphere head-related transfer functions (RS-HRTF). Performance was measured using inter-zone isolation (IZI), inter-program interference (IPI), and crosstalk cancellation (XTC) across 100-20,000 Hz. The results provide actionable engineering insights: FR acts as a spectral calibrator, DIR delivers the most reliable gains for creating separate sound zones (10.05 dB average improvement in IZI/IPI), and RS-HRTF is critical for binaural separation, boosting XTC from an average of 4.51 dB to 7.91 dB, with effects strongest above 2 kHz. This quantified breakdown directly guides audio engineers on where to invest limited resources—be it measurement time or computational cost—when constructing training datasets for neural audio rendering systems, prioritizing components with the highest impact on real-world performance.