Research map / 3D vision / 3D generation
Resounding Acoustic Fields with Reciprocity
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes resounding, estimating room impulse responses at arbitrary emitter positions from sparse measured emitters, using acoustic reciprocity via Versa-ELE data augmentation and Versa-SSL self-supervision on top of existing neural acoustic field models.
Problem
Estimating room impulse responses at arbitrary emitter locations when only a sparse set of emitter positions is measured (dense listener positions per emitter), i.e., the resounding task analogous to relighting.
Approach
Exploit acoustic reciprocity by swapping emitter and listener roles to create physically valid extra training pairs (Versa-ELE), and enforce reciprocity via a self-supervised consistency loss with stop-gradient on one branch using gain patterns from a pretrained AVR model (Versa-SSL); designed as a plug-in augmentation to existing neural acoustic field methods.
Key results
- On AcoustiX, 10x fewer listener positions with ELE improves AVR's T60 error from 15.1 to 12.6 and STFT from 1.9 to 1.3 (per rebuttal table).
- Versa-SSL with ground-truth gain pattern slightly improves over learned gain pattern on Acoustic-Diff (STFT 1.39 to 1.31, EDT 21.5 to 20.6).
- Versa consistently improves NAF, INRAS, and AVR across 5 to 25 emitter settings on MeshRIR and AcoustiX (per rebuttal figures).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ELE to AVR (ablation) | AVR · GWA | STFT: 2.2 -> 1.7 | Helped |
| Add ELE to AVR (ablation) | AVR · GWA | Amp.: 0.82 -> 0.57 | Helped |
| Add ELE to AVR (ablation) | AVR · GWA | Env.: 11.1 -> 7.3 | Helped |
| Add ELE to AVR (ablation) | AVR · GWA | T60: 22.9 -> 16.8 | Helped |
| Add ELE to AVR (ablation) | AVR · GWA | C50: 2.9 -> 2.1 | Helped |
| Add ELE to AVR (ablation) | AVR · GWA | EDT: 45.4 -> 27.9 | Helped |
| Use GT gain pattern instead of learned gain pattern (ablation) | Versa-SSL · Acoustic-Diff | STFT: 1.39 -> 1.31 | Helped |
| Use GT gain pattern instead of learned gain pattern (ablation) | Versa-SSL · Acoustic-Diff | Amp.: 0.25 -> 0.24 | Helped |
| Use GT gain pattern instead of learned gain pattern (ablation) | Versa-SSL · Acoustic-Diff | Env.: 3.6 -> 3.4 | Helped |
| Use GT gain pattern instead of learned gain pattern (ablation) | Versa-SSL · Acoustic-Diff | T60: 15.7 -> 15.5 | Helped |
| Use GT gain pattern instead of learned gain pattern (ablation) | Versa-SSL · Acoustic-Diff | C50: 2.1 -> 1.9 | Helped |
| Use GT gain pattern instead of learned gain pattern (ablation) | Versa-SSL · Acoustic-Diff | EDT: 21.5 -> 20.6 | Helped |
Surprising findings
- Using the learned gain pattern from AVR rather than ground truth only slightly degrades performance, suggesting the method is robust to gain-pattern inaccuracies.
Limitations
- Assumes geometric (ray-based) acoustics and linear time-invariant propagation; low-frequency wave effects are not modeled.
- Versa-SSL requires position and orientation information and depends on a pretrained AVR model for gain patterns.
- Roughly doubles training cost for both ELE and SSL variants.
Experimental setup
Datasets: MeshRIR AcoustiX GWA SoundSpaces (excluded)
Benchmarks: MeshRIR AcoustiX GWA
Metrics: STFT distance amplitude error energy decay (Env.) T60 C50 EDT
Baselines: NN Linear NAF INRAS AV-NeRF AVR
Builds on: neural acoustic fields AVR NAF INRAS
Closest papers by meaning
- Hearing Anything Anywhere (CVPR 2024 · similarity 0.88)
- NeRAF: 3D Scene Infused Neural Radiance and Acoustic Fields (ICLR 2025 · similarity 0.88)
- Deep Neural Room Acoustics Primitive (ICML 2024 · similarity 0.88)
- SoundVista: Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding (CVPR 2025 · similarity 0.86)
- Acoustic Volume Rendering for Neural Impulse Response Fields (NeurIPS 2024 · similarity 0.86)
- Hearing Anywhere in Any Environment (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.