Binaural Scene Transfer: A Deep Learning Framework for Cross-Environment 3D Audio Rendering on Stereo Headphones

This paper introduces Binaural Scene Transfer (BST), a software-only framework that enables immersive 3D audio experiences using standard stereo headphones, without requiring head tracking or specialized playback systems. The proposed method leverages U-Net, a deep generative model, to map the spatial and reverberant characteristics of various environments-such as concert halls, forests, or cathedrals-onto dry mono or stereo audio in real time. The system comprises four key components: input and environment conditioning; a scene-conditioned audio rendering model trained on measured and simulated BRIRs; an HRTF-aware spatial audio encoder that personalizes binaural cues; and a web-based front-end backed by a Python inference server for interactive demos and listening tests. Key features of the framework include AI-based 3D spatialization from stereo input/output alone, compatibility with any standard headphones, environment-specific acoustic transfer applied to arbitrary audio, and optional integration with game audio pipelines or digital audio workstations (e.g., Unreal Engine, Unity, Reaper) as a plugin.