Research map / 3D vision / NeRF and radiance fields
Acoustic Volume Rendering for Neural Impulse Response Fields
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 Acoustic Volume Rendering (AVR), which renders impulse responses in the frequency domain with spherical integration to learn neural impulse response fields, plus AcoustiX, a physics-based acoustic ray-tracing simulator.
Problem
Learning neural impulse response fields that synthesize room impulse responses at novel emitter/listener poses; existing neural field methods overfit and fail to capture waveform shape, phase, and time-of-arrival, and existing simulators have phase and timing inaccuracies.
Approach
Adapts NeRF-style volume rendering to time-series impulse responses by rendering in the frequency domain (time delays become phase shifts), integrating over a sphere of rays with a listener gain pattern, and supervising spectral, phase, amplitude, and time-domain losses.
Key results
- On MeshRIR (0.1s IR), AVR reduces phase error to 0.85 from 1.61 for NAF and INRAS and envelope error to 1.15 vs 1.59 for NAF.
- On RAF-Furnished (0.1s IR), AVR achieves amplitude error 0.75 and T60 error 5.0, outperforming NAF (0.93, 7.1).
- In a 7-user study of zero-shot binaural rendering, AVR scored 4.71 out of 5 versus 1.42 for NAF and 1.86 for INRAS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use acoustic volume rendering instead of the strongest baseline | AVR · MeshRIR | Phase error: 1.43 -> 0.85 | Helped |
| Use acoustic volume rendering instead of the strongest baseline | AVR · RAF-Furnished | Amplitude error: 0.93 -> 0.75 | Helped |
| Use acoustic volume rendering instead of the strongest baseline | AVR · RAF-Furnished | Envelope error: 3.81 -> 4.52 | Hurt |
| Use acoustic volume rendering instead of the strongest baseline | AVR · RAF-Empty | Phase error: 1.59 -> 1.58 | Helped |
| Use acoustic volume rendering instead of the strongest baseline | AVR · 2D Room | Envelope error: 2.53 -> 2.35 | Helped |
| Use acoustic volume rendering instead of the strongest baseline | AVR · 3D scene Avonia | Amplitude error: 0.71 -> 0.63 | Helped |
| Use acoustic volume rendering instead of the strongest baseline | AVR · 3D scene Montreal | Envelope error: 5.19 -> 5.16 | Helped |
| Replace time-domain rendering with frequency-domain rendering (ablation) | AVR · MeshRIR | Phase error: 1.181 -> 0.847 | Helped |
| Add raw signal loss (ablation) | AVR · MeshRIR | Amplitude error: 0.558 -> 0.535 | Helped |
| Add raw signal loss (ablation) | AVR · MeshRIR | EDT: 46.4 -> 35.1 | Helped |
| Add angle and spectral loss (ablation) | AVR · MeshRIR | Phase error: 1.453 -> 0.847 | Helped |
| Add angle and spectral loss (ablation) | AVR · MeshRIR | C50: 2.65 -> 0.92 | Helped |
Surprising findings
- Existing learning-based methods (NAF, INRAS) fail to learn valid phase information, with phase error at the random-phase level (1.62).
- AVR inference is roughly 10x slower than NAF and INRAS (30.3 ms vs 3.2 ms and 2.1 ms at 0.1s IR).
Limitations
- Rendering samples many rays and points per query, causing high memory use and slower inference.
- A new model must be trained per scene, requiring impulse response measurements in each new room.
- Zero-shot binaural evaluation is limited to a single user study with seven participants.
Experimental setup
Datasets: MeshRIR Real Acoustic Field (RAF) AcoustiX simulated 2D room AcoustiX simulated 3D scenes Avonia and Montreal from iGibson
Benchmarks: MeshRIR RAF-Furnished RAF-Empty AcoustiX 2D Room AcoustiX 3D Avonia AcoustiX 3D Montreal
Metrics: phase error amplitude error envelope error T60 C50 EDT
Baselines: NAF INRAS AAC Opus AV-NeRF
Builds on: NeRF NAF INRAS Sionna
Closest papers by meaning
- Hearing Anything Anywhere (CVPR 2024 · similarity 0.88)
- AV-GS: Learning Material and Geometry Aware Priors for Novel View Acoustic Synthesis (NeurIPS 2024 · similarity 0.87)
- Real Acoustic Fields: An Audio-Visual Room Acoustics Dataset and Benchmark (CVPR 2024 · similarity 0.87)
- Hearing Anywhere in Any Environment (CVPR 2025 · similarity 0.86)
- Resounding Acoustic Fields with Reciprocity (NeurIPS 2025 · similarity 0.86)
- NeRAF: 3D Scene Infused Neural Radiance and Acoustic Fields (ICLR 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.