AES Show 2024 NY

To achieve highly realistic 3D audio reproduction in virtual reality (VR) or augmented reality (AR) through binaural rendering, we must address the considerable computational complexity involved in convoluting head-related impulse responses (HRIRs). To reduce this complexity, an algorithm is proposed where audio signals are distributed to pre-defined representative directions through panning. Only the distributed signals are then convoluted with the corresponding HRIRs. In this study, we explored a method for generating representative HRIRs through learning, utilizing a full-sphere HRIR set. This approach takes into account smooth transitions and minimal degradation introduced during rendering, for both moving and static audio objects. Compared with conventional panning, the proposed method reduces average distortion by approximately 47% while maintaining the runtime complexity of the rendering.