AMBIT generates and edits first-order ambisonics with executable plans
Original title: Spatial Audio Generation and Editing with Executable Plans and Self-Distilling Transfusion
A shared Transfusion model plans scenes, renders FOA audio, and self-distills on student prefixes.
AMBIT is a framework for native first-order Ambisonics generation and editing. Executable ScenePlans specify source content, activity, and 3D trajectories, while an FOA-aware continuous VAE preserves content and inter-channel structure. Editing uses contrastive features for the requested change and reference latents for residual acoustic detail. On-policy self-distillation refines planning and rendering. Evaluations on speech, music, and sound report better content, spatial quality, editing fidelity, and preservation than the compared baselines.
Unconfirmed: Quality gains are stated relative to unnamed compared baselines, without numeric scores in the text.
- Plans describe the full target scene, including sources that should be retained.
- The VAE uses shared-decoder recovery, asymmetric grouped KL, and spatial covariance matching.
- Target-conditioned teachers supervise both modes of the shared student.
- A project page is published at wjm-george.github.io/AMBIT.
From the original article
1 min read at huggingface.co
Spatial audio authoring coordinates source decisions with soundfield synthesis. For editing, plans specify what to change and retain, while reference audio supplies further acoustic detail. We present AMBIT, a unified framework for native First-Order Ambisonics (FOA) generation and editing with three connected contributions.
By Kam Man Wu, Yu Zhang, Pengjun Fang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.