acceptodds
Under review as a conference paper at ICLR 2027

Floor-Plan-Conditioned Generative Priors for Sparse Sound-Field Reconstruction

Abstract

Many spatial audio applications, such as rendering AR/VR acoustics or room correction, require knowledge of the sound field throughout a room, yet measuring it densely takes hundreds of microphone positions. With only a few microphones, many different fields fit the same measurements. Floor plans are often available and constrain the field before any measurement is taken. However, floor plans describe only the walls and doorways, not the furniture and materials that also shape the acoustic field. We therefore learn a distribution of sound fields consistent with a floor plan, using a flow-based generative prior that maps a cheap image-source field, computed from the floor plan alone, to high-fidelity wave-simulated fields. Given a few measurements, posterior sampling recovers the field of a specific room. Because the model is generative, the variance across posterior samples gives an uncertainty map along with the reconstruction. This map tracks the reconstruction error, so it can indicate where to place the next microphones. We evaluate on real multi-room apartment floor plans with wave-simulated acoustics. On floor plans unseen during training, our method reduces average reconstruction error by 5.5dB with 128 microphones relative to kernel interpolation given the same floor-plan information, and the gain grows as microphones are added. With 16 microphones, it is at least as accurate as the same baseline with 64. Active sampling based on posterior uncertainty improves reconstruction by up to 0.7dB over spatially uniform placement, reaching the same accuracy with 15% fewer microphones.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Generative Priors, Sound-Field Reconstruction, Posterior Sampling, Active Sensing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_4QgfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a (weakly) geometry-informed sound field estimation over a wide receiver region under an extremely sparse measurement setup. Specifically, a flow-matching model is trained to serve as a generative prior of a sound field under a given floor plan and source position. The uncertainty it learns here is the variance caused by unknown indoor objects (e.g., furniture), not captured in the given floor plan. Then, for the sound field estimation, DPS is employed with a "reconstruction guidance" given by the sparse measurements. The proposed model was trained and evaluated on synthetic indoor scenes + a wave-based solver. Comparisons were made with kernel interpolation methods and recent neural acoustic field models. Further, the authors explored finding the next plausible measurement positions (i.e., active sampling) using the posterior variance and showed that it has benefits over the other simple heuristic.

Critical strengths

Overall, the authors proposed a novel idea, introducing geometry-informed generative priors for improved sound field reconstruction and active sampling strategies. The ablation study was particularly interestin for me, and it justifies many design choices presented in the paper.

Critical weaknesses

I find the baseline comparison part a bit weak.

  1. No evaluation on real-world measurements was conducted.

  2. It is good to see that the proposed model outperforms AVR and INFER. Still, they come from the "acoustic field learning" literature, which has a slightly different focus; they are developed for full-band response estimation, do not always include geometry information, and are not optimized for NMSE (so there is no surprise that AVR and INFER fail to outperform the basic kernel interpolation on that metric).

  3. The comparison with kernel interpolation methods is also somewhat weak. The "matched kernel" version helps a bit as it incorporates the ISM field, but it is still a simple kernel interpolation that is unaware of the source position or floor plan. I find the following more suitable as baselines, they are also learned [1, 3], geometry-informed [1, 2], and use NMSE as their main metric [1, 2, 3].

    [1] Damiano, Stefano, and Toon Van Waterschoot. "Sound field reconstruction using physics-informed boundary integral networks." EUSIPCO, 2025.

    [2] Sundström, David, Filip Elvander, and Andreas Jakobsson. "Boundary-informed sound field reconstruction." EUSIPCO, 2025.

    [3] Ribeiro, Juliano GC, et al. "Sound field estimation based on physics-constrained kernel interpolation adapted to environment." IEEE/ACM TASLP, 2024.

The trained system also seems very sensitive to errors and uncertainties (beyond the interior objects being modeled), including the floor plan error, as the author noted, and possibly to source/measurement-position errors.

Questions for authors

Please correct me if I am wrong in any of these questions and comments.

What is the dimensionality of the sound field in question, 2D or 3D? The paper reads like the former is the case (as the appendix says that the ISM uses a zero-order Hankel function), but it would be nice to state it more upfront.

Do the authors think the sensitivity to the geometry error can be potentially mitigated, e.g., with data augmentation? Or is it a fundamental limitation of the proposed method? I suspect it is a mix of both, but can we quantify each contribution? I find this question particularly important because of its practical relevance.

Given that the paper tackles an extremely wide region of a sound field, I am unsure whether (the current definition of) NMSE is appropriate for evaluation. Since its denominator is a sum over the entire field, it does not normalize the local loudness and downweights the quiet region (I believe this problem does not appear in most sound field estimation work because they are concerned with small local receiver area/volume).

In a similar vein, as shown in Figure 4, the posterior variance seems to be mostly governed by the magnitude of the sound field itself, and as a result, the proposed method places most of the receivers at the loud regions, resulting in higher errors in quiet regions than the simple farthest-point placements (Figure 3). In other words, the current framework seems to downweight the correctness and informativeness of quiet regions (informally speaking), which the current NMSE metric does not reflect appropriately.

Can we also visualize the total posterior variance reduction Δ(ρ)\Delta (\rho) in Figure 4?

I am unsure how practical the current source placement method is. It requires repeated alternation between measurements and DPS inference, and this must be done for every source position.

Can we get a plot of errors across different frequency bands, as included in most sound field estimation papers? Also, does the placement preference (given by the posterior uncertainty) change across different frequency bands?

Reviewer_Ue6LOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a floor-plan-conditioned flow prior for sparse sound-field reconstruction. Measurement-guided sampling produces the reconstructed field and an uncertainty map, which guides additional microphone placement.AI was used only for language polishing. All assessments are my own.

Critical strengths
  1. The framework addresses a meaningful problem and combines reconstruction with active measurement selection.
  2. Reconstruction gains are substantial, including against a kernel baseline using the same ISM information.
  3. The experiments include useful controls on noise, training seeds, floor-plan errors, and microphone placement.
Critical weaknesses
  1. The main results use gain and delay calibrated from the complete simulated empty-room field (Appendix C.2). Microphone-only calibration gives similar reconstruction accuracy at larger budgets, but active sampling should also be validated under this deployable protocol.
  2. The uncertainty evaluation demonstrates error ranking rather than calibration. The correlation in Section 4.3 does not establish whether the approximate posterior samples provide reliable credible regions.
  3. Evaluation is limited to two-dimensional simulations at 42–600 Hz (Section 4.1). Real-room generalization remains unverified, and Table 7 shows sensitivity to systematic floor-plan errors.
  4. The geometric contribution is not fully isolated. Table 2 replaces the ISM field but retains the floor-plan mask. Separate ablations and a geometry-aware baseline would clarify the benefit of each input.
Questions for authors
  1. Do the active-sampling gains remain when calibration uses only the currently available microphones, starting from four measurements?
  2. Can the authors evaluate uncertainty calibration in addition to rank correlation?
Reviewer_fWgFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a floor-plan-conditioned generative prior for reconstructing sound fields from sparse microphone measurements. It conditions on a 2D image of the sound field using ISM with the floor plan, and reconstructs a room through posterior sampling with additional loss that measures the mismatch with sparse microphone measurements. Posterior uncertainty also guides the placement of additional microphones. Experiments show improved reconstruction over kernel interpolation and two baseline methods (INFER and VAR).

Critical strengths
  1. The generative prior and guided posterior sampling effectively combine prior knowledge with measurements.

  2. Posterior samples provide spatial uncertainty estimates that guide where to measure next (active sampling).

Critical weaknesses

1. Generalization to real measurements is not demonstrated.

While the floor plans are a real-world dataset, the acoustic measurements are all simulated. It also remains unclear how well the system would perform in practice.

2. The 2D setting omits important room acoustics.

The simulations do not model source or microphone height, floor/ceiling reflections, or propagation over and under furniture. These effects matter even when microphones lie on a single horizontal plane.

3. The frequency range tested is very narrow.

Evaluation covers only 42–600 Hz, while prior methods like NAF, INFER, and AVR cover much larger frequency range. This small range is relevant to low-frequency room acoustics, but does not establish broadband speech, music, or AR/VR rendering.

4. The failure of INFER and AVR needs a clearer explanation.

INFER and AIR doesn't seem to work for the setting tested in the paper. Both method have positive average NMSE at every microphone budget, which is worse than a trivial zero-field prediction (NMSE=0dB). AVR also becomes worse as microphones are added. The authors should explain more clearly why those baselines methods are completely failing here.

5. The evaluation is difficult to connect to established acoustic benchmarks.

The paper primarily reports NMSE on its own simulated dataset. Prior work also evaluates spectral amplitude and phase errors, impulse-response envelopes, reverberation time (T60), early decay time (EDT), and clarity (C50), including on real-world datasets such as MeshRIR and RAF. Changing the data, bandwidth, dimensionality, and metrics together makes it difficult to assess whether the reported gains translate into better acoustic rendering.

Questions for authors
  1. How does the proposed method perform real-world measurements or datasets? Do the reconstruction gains and the reliability of uncertainty-guided sampling survive the simulation-to-real gap?

  2. Why is this method limited to 42–600 Hz? Any evidence suggests that the microphone savings extend to higher frequencies?

  3. What are the main challenges for adapting this method in 3D?

  4. Why do INFER and AVR perform worse than a zero-field prediction? Why does AVR worsen with more microphones?

  5. How does the proposed method perform in terms of T60, EDT, and C50?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.