Headworn microphone arrays are increasingly used for spatial audio, but their small aperture limits the resolution of direction-of-arrival (DOA) estimation, especially in reverberant environments. Signal -dependent binaural reproduction methods rely on estimating DOA and applying a corresponding directional filter or mapping, but this approach degrades under low spatial resolution and dynamic scenes. We propose a performance-weighted blended (PWB) approach to binaural reproduction that avoids explicit DOA estimation. Instead of selecting a single directional beamformer, we construct a bank of candidate beamformers corresponding to multiple spatial hypotheses and blend their outputs using data-driven weights. These weights are computed from the residual error in approximating the received signals, enabling signal-dependent adaptation that emphasizes the most perceptually relevant contributors. We evaluate the proposed method using acoustic scenes with head-worn arrays, varying reverberation, source positions, and motion trajectories. Our analysis includes interaural time differences (ITDs), interaural level differences (ILDs), and overall reconstruction error. Comparisons against existing techniques—including directional binaural signal matching and COMPASS—demonstrate that PWB reconstruction accurately preserves spatial cues and maintains robustness in challenging conditions. We show that residual-based blending provides perceptually consistent binaural rendering without requiring accurate localization.
Mittal et al. (2025) studied this question.