acceptodds
Withdrawn from ICLR 2027

CERPA: Heading-Equivariant Attention for Yaw-Invariant Few-Shot Room-Acoustic Synthesis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

State-of-the-art cross-room few-shot acoustic synthesis methods generate room impulse responses (RIRs) from sparse acoustic context and geometry conditioningf. However, their geometry conditioning often fails to respect the heading invariance of monaural acoustics: jointly applying a yaw rotation to the panoramic depth image and source–receiver poses can change the predicted RIRs. Such a change of heading reference changes only the coordinate frame, leaving the underlying acoustic characteristics unchanged; the conditional probabilistic RIR distribution or deterministic RIR output should therefore remain invariant. To address such physical inconsistency, we introduce CERPA (Cyclic-Equivariant Relative Positional Attention), a unified attention framework that explicitly builds this rotational invariance into the visual representation. CERPA modifies the attention mechanism of the geometry-conditioning ViT to produce cyclic rotation equivariant tokens, followed by an invariant readout for heading-invariant visual conditioning. When integrated into both deterministic and state-of-the-art probabilistic acoustic models, CERPA ensures yaw invariant acoustic synthesis. In both simulated and real-world few-shot RIR synthesis, our approach substantially improves heading consistency and outperforms non-equivariant baselines on key acoustic metrics.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
geometric deep learning, group equivariance, relative positional encoding, vision transformers, panoramic geometry, few-shot acoustic synthesis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_SrBLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CERPA, a heading-equivariant attention framework for yaw-invariant few-shot room-acoustic synthesis. It introduces circular relative-position biases and periodic rotary embeddings to ensure equivariant panoramic representations and invariant RIR predictions. The method is integrated into xRIR and FLAC, demonstrating improved heading consistency and competitive acoustic synthesis performance on simulated and real-world datasets.

Critical strengths
  1. Well-motivated architectural design. The paper identifies an important physical consistency issue in room-acoustic synthesis and provides a mathematically grounded solution with architectural equivariance guarantees.

  2. Validation across different frameworks. CERPA is evaluated on both deterministic (xRIR) and probabilistic (FLAC) acoustic synthesis models, demonstrating significant improvements in heading consistency. The parameter-efficiency results are also encouraging.

Critical weaknesses
  1. Limited methodological novelty over existing equivariant positional encodings. The core idea of enforcing cyclic shift equivariance through relative positional encoding has been extensively explored. In particular, Making Vision Transformers Truly Shift-Equivariant (CVPR 2024) already establishes circular shift-equivariant transformer architectures, while Algebraic Positional Encodings (NeurIPS 2024) provides a general formulation for positional representations over finite cyclic groups. More recently, SpheRoPE (2026) also employs harmonically constrained rotary embeddings to enforce periodicity in panoramic representations. CERPA appears to adapt these established principles to room-acoustic synthesis through cylindrical coordinate transformations and invariant pooling. While this application is meaningful, the distinct algorithmic contribution remains unclear. Moreover, the paper lacks sufficient ablations to isolate the contributions of cylindrical preprocessing, relative positional encoding, invariant pooling, and pose conditioning.

  2. Limited generalization of the invariance guarantees and inconsistent improvements in acoustic quality. The theoretical guarantees and consistency experiments focus on discrete, patch-aligned yaw rotations rather than arbitrary continuous rotations. It remains unclear whether the method maintains similar robustness under realistic off-grid rotations. Furthermore, improved heading consistency does not consistently translate into better acoustic synthesis. In Table 1, CylindricalViT performs worse than the original xRIR across all six metrics under K=1. In Table 4, CylDINO also underperforms the original FLAC on most real-world metrics under K=8, including T60, EDT, and retrieval recall. These results weaken the claim that enforcing heading invariance consistently improves acoustic synthesis quality.

Questions for authors

1.Novelty: What are the key technical differences between CERPA and existing cyclic-equivariant transformers or periodic rotary positional encodings, such as SpheRoPE? Could the authors provide component-wise ablations to clarify the novel contributions?

2.Continuous rotation: How does CERPA perform under arbitrary yaw angles that are not aligned with the patch grid? Can the authors evaluate both invariance error and RIR reconstruction quality under these rotations?

3.Acoustic performance: Why does CERPA degrade performance in several settings, particularly xRIR with K=1 and FLAC on the real-world HAA dataset? Could the authors clarify the trade-off between strict heading invariance and acoustic reconstruction accuracy?

Reviewer_bQsuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies a few-shot acoustic field learning task. The authors note that the omnidirectional receiver's responses are invariant to yaw, yet existing methods do not exploit this. Accordingly, the authors propose a modification to a ViT backbone (used in recent prior work) that respects this symmetry. This is achieved by equivariance in attention and invariance in readout. Two variants are presented, one with additive bias (for xRIR) and the other with a custom RoPE (for FLAC). The experimental setup follows prior work.

Critical strengths

Enforcing the yaw invariance itself is an interesting idea, which shows a bit of performance gain.

Critical weaknesses

The main results show that the yaw augmentation is sufficient to achieve essentially the same performance, which undermines the contribution significantly. This is not so surprising, because regarding the yaw invariance, each input’s orbit is one-dimensional and can be densely sampled with augmentation. Any strong enough unconstrained model will learn that fairly well.

Questions for authors
  1. Could the authors provide results of the yaw-augmented baseline for Tables 2 and 3?
  2. It seems like the proposed method works only for omnidirectional receivers. Can it be extended for, e.g., binaural or spatial RIRs?
Reviewer_bDe5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses heading dependence in few-shot room impulse response (RIR) synthesis. Jointly rotating the panoramic depth map and the source and receiver poses changes only the coordinate frame, so for an omnidirectional source the predicted monaural RIR should not change, yet a released FLAC checkpoint changes its prediction by about 20 percent in relative L2 under such rotations (App. C.1). The paper proposes CERPA, attention with positional terms depending only on elevation and circular azimuth offsets. This makes it equivariant to cyclic shifts of the panorama by whole patch columns. An additive version, CylindricalViT, is placed in xRIR and a rotary version with integer harmonics in FLAC, followed by mean pooling and a heading-invariant pose encoding. On AcousticRooms both variants are almost exactly invariant under patch-aligned rotations, with acoustic metrics comparable to the hosts and also to yaw-augmented baselines. The paper also reports parameter and data efficiency studies and fine-tuning on the real HAA rooms.

Critical strengths

The symmetry is stated precisely, and the construction that enforces it seems correct and simple tooo. Eq. 1 defines what should be invariant and which inputs move under the group action, Proposition 1 follows from the invariance of the relative descriptor, and the integer-harmonic rotary encoding (Eq. 7/8) handles phase closure on the cylinder. The same works in a deterministic model trained from scratch and in a pretrained DINOv3 backbone feeding a flow-matching generator, and Fig. 7 verifies equivariance at the encoder as well as invariance at output. This matters for the assessment because it turns heading consistency into a property of the architecture rather than something hoped for from augmentation.

The pilot study in Appendix C.1 shows that the problem is real in a published model and that standard metrics can hide it. With the checkpoint, noise and settings fixed, a pure change of heading changes the FLAC prediction by 19 to 22 percent in relative L2 and the EDT by about 19 to 20 ms, while the aggregate EDT error moves by only 3.5 to 6.3 ms (Table 7). This is the most informative result in the paper and the main reason the work is of interest beyond the specific architecture. It would fit well in the main text.

Critical weaknesses

However, evidence for better acoustic quality does not yet separate the method from the baseline and training variance. Released FLAC checkpoint evaluated at 0 degrees in Table 7a, on the same unseen split with one context RIR, does better than both the retrained FLAC and FLAC with CylDINO in Table 1. Each variant is trained once and the reported spread covers evaluation seeds only, while the xRIR baseline appears with C50 1.238 in Table 1 and 1.275 in Table 2a, gap of about five times the C50 advantage of CylindricalViT. Further, against the yaw-augmented baselines, the results seem mixed. CylDINO is worse on C50 and T60 at both context sizes. On HAA, only the CERPA variants receive a loudspeaker orientation input (App. C4/C5), and CylDINO still trails FLAC on 5 metrics with 8 context RIRs. In my opinion, this matters given the improvement claims in the abstract and Conclusion are what would carry the paper beyond a known construction. Further, Table 1 supports comparable quality with guaranteed invariance.

This guarantee seems to hold for 16 or 32 patch-aligned rotations, and it's evaluated only there. CylindricalViT is equivariant in steps of 22.5 degrees and CylDINO in steps of 11.25 degrees, while the yaw-augmented baselines are trained on all 512 pixel-aligned shifts. Figures 3, 7 test only patch-aligned angles, where the error is zero. I think they confirm implementation more than benefit. Exact invariance over a group this small can also be available without changing the architecture, by averaging an unmodified encoder over the N shifts (Puny et al., ICLR 2022) or by canonicalising the heading (Kaba et al., ICML 2023). Table 5 notes that frame averaging was not used, but it is not compared. I think this matters because the practical case for the architecture over these simpler routes is not yet shown, and the title and abstract claim yaw invariance without the patch-aligned qualifier, used by Section 3. An off the grid evaluation and this comparison should be able to address it.

Questions for authors
  • Are the Table 7a results for the released FLAC checkpoint at 0 degrees directly comparable with the FLAC row of Table 1?
  • What training budget and encoder size does the released checkpoint use? Moreover, how does CylDINO compare when trained to the same budget?
  • On HAA, how do the baselines perform when they receive the same loudspeaker orientation input as the CERPA variants?
  • What is the prediction invariance error of each model at pixel shifts that are not multiples of the patch width?
  • Further, how do the metrics in Table 1 change when test headings are drawn randomly? I think this is the setting where the architectural guarantee should matter.

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

Related papers

Discussion (0)

Sign in to comment.