Reproduce “AudioSphere: Towards Self-Supervised Spatial Audio Representation Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Goksenin Yuksel, Marcel van Gerven, Kiki van der Heijden

AudioSphere learns a shared audio-spatial representation by masking multichannel sound and reconstructing acoustic patterns and directional cues. Its frozen features transfer across recognition and spatial event tasks. On TAU-2019, full-model localization error is 14.9°, versus 73.0° to 73.5° for reduced-input variants, supporting a role for directional information. On TAU-2021, adding a temporal head raises frozen-probe F-score from 21.0 to 37.2; the shallow probe alone does not establish moving-scene tracking.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.