BiFormer3D: Grid-Free Time-Domain Reconstruction of Head-Related Impulse Responses with a Spatially Encoded Transformer

TL;DR

BiFormer3D is a novel time-domain, grid-free Transformer model that reconstructs head-related impulse responses at arbitrary directions from sparse measurements, improving spatial and temporal fidelity.

Contribution

It introduces a time-domain, grid-free Transformer architecture with sinusoidal spatial features for HRIR up-sampling, surpassing prior frequency-domain methods.

Findings

Improves NMSE, cosine distance, and ITD/ILD errors on SONICOM dataset.

Ablation studies validate the effectiveness of proposed modules.

Minimum-phase pre-processing is unnecessary for the model.

Abstract

Individualized head-related impulse responses (HRIRs) enable binaural rendering, but dense per-listener measurements are costly. We address HRIR spatial up-sampling from sparse per-listener measurements: given a few measured HRIRs for a listener, predict HRIRs at unmeasured target directions. Prior learning methods often work in the frequency domain, rely on minimum-phase assumptions or separate timing models, and use a fixed direction grid, which can degrade temporal fidelity and spatial continuity. We propose BiFormer3D, a time-domain, grid-free binaural Transformer for reconstructing HRIRs at arbitrary directions from sparse inputs. It uses sinusoidal spatial features, a Conv1D refinement module, and auxiliary interaural time difference (ITD) and interaural level difference (ILD) heads. On SONICOM, it improves normalized mean squared error (NMSE), cosine distance, and ITD/ILD errors over prior methods; ablations validate modules and show minimum-phase pre-processing is unnecessary.