Resounding Acoustic Fields with Reciprocity

We thank Reviewer sQL3 for the careful review. We appreciate the reviewer for recognizing that our method shows “consistent performance gains for different configurations”, and our used datasets “cover diverse types of acoustic observations”. We address your concerns as follows.

---

Review W1 part1: Question on Versa on Simulated dataset.

In the simulated environment, one can freely simulate the IRs at any position. The reason why we incorporate the simulated dataset in this paper is to demonstrate the effectiveness of our method in various dataset settings. We answer the rest of the questions (W1 part2, W2 and Q1) about whether reciprocity holds in the real world environment as follows.

---

Review Q1(1), W2, W1 part2: Whether requirements for reciprocity are met in the real dataset?

The Rayleigh reciprocity theorem assumes linear, time-invariant (LTI) systems and symmetric propagation media. In a simulated environment, these could be easily satisfied through ray-based or wave-based methods. In real environment, most room acoustic environments used for impulse response measurements (e.g., MeshRIR, RAF) meet these criteria:

1. The environment is static during measurement and there is no moving object or human in the scene, making it time-invariant. (As also noted in the dataset collection procedure in the MeshRIR and RAF paper)
2. All common materials and structures (e.g., drywall, wood, concrete, furniture, carpeted floors) exhibit symmetric acoustic reflection behaviors. Only specially-engineered materials like metamaterials with time-varying impedance, active acoustic diodes, or isolators exhibit non-symmetric reflections. However, these materials are mostly found only in laboratories or engineered acoustic environments, and are not presented in daily life, especially in room acoustic measurement scenes.

Based on the above analysis, real datasets and real-life scenarios would meet the requirements for reciprocity. We also show some RIR measurements in real-world datasets to prove reciprocity in the following answers.

---

Review Q1(2), W2, W1 part2: Preservation of IRs after position swaps in real dataset.

It is easy to verify the reciprocity in the simulated dataset. For this question, we focus on proving reciprocity on the real dataset.

Discussion on MeshRIR and RAF:

MeshRIR has different areas for emitter/listener positions, and thus does not include RIR pairs that have the same emitter/listener positions after swapping. RAF has significantly different emitter and receiver gain patterns, nor do they have listener orientation measurements. Thus, there isn’t any RIR pair that have the same emitter/listener positions and orientations after swapping. As a result, we could not empirically evaluate the impulse response invariance for these two datasets based on their current data points. However, experimental results in our original paper show consistent improvements of Versa on them, indicating reciprocity constraints do hold for them.

Discussion on our collected real dataset as a proof:

Though we can not get the pairs of IRs after switching the position of the listener/emitter in the MeshRIR and RAF dataset (these samples are not included in the dataset), we collect our own real dataset as proof.

Dataset collection procedure: We conducted an experiment in 3 various indoor environments (an office room(9m x 8m, a conference room (13m x 7m), a kitchen (5m x 9m)) using a microphone (ADMP401 MEMS microphone) and a speaker (Bose SoundLink Revolve II speaker) that are both omnidirectional. They are wired connected to an audio interface (BELA board), which can synchronize between the playing and recording. We follow the data collection in DiffRIR and play a 10-second chirp signal that spans 50Hz to 20KHz to collect RIR samples. We collected the impulse responses emitted from position A to B and from position B to A. Under the reciprocity principle, these pairs of impulse responses should be identical. Positions A and B are randomly chosen to represent the diverse acoustic fingerprint in an environment. We collect 20 such RIR pairs in each environment, placing them in a diverse combination of speaker and microphone positions.

Results: We measure the similarities within pairs of RIRs that have swapped emitter/listener positions and the cross-paired RIRs. As shown in the table, in each of the scenes, the paired RIRs have much lower acoustic metrics (some metrics are x10 smaller) compared with cross-paired RIRs. This confirms that the RIRs are preserved in terms of the position swaps for various real-world environments.

We will also incorporate these supplementary results in the revised version to show the reciprocity for the real-world environment.

---

Review Q1(3), W2, W1 part2: Provide evidence of NN have learn reciprocity with Versa.

We acknowledge the importance of validating whether the neural acoustic field models trained with Versa indeed exhibit reciprocity. We included a similar experiment in Q1(2). We randomly query the trained AVR and NAF models on swapped emitter/listener pairs at each scene and evaluate the predicted impulse response similarity. As shown in the table, the predicted RIRs pairs with Versa training exhibit much lower acoustic metrics (showing higher similarities for paired predictions after position swap) compared with the neural acoustic field without the Versa training. These show that the NN models learn reciprocity through the Versa strategy. In addition, we promise to include more waveform visualizations of the model-predicted RIRs in the final manuscript.

---

Review Q2: Questions on the perceptual user study.

• IRB and Recruitment: We have consulted with our host institution's IRB office regarding the ethical requirements of our audio listening evaluation. Based on the nature of the study that involves only anonymous adult participants, no collection of personal or identifiable data, and no procedures involving risk or discomfort, the study was formally deemed exempt from IRB review and granted a waiver with “non-human subject determination”. We will clarify this waiver status in the revised paper to avoid confusion.

• Background: Though the participants were not expert acousticians, we ensured they had sufficient familiarity with spatial audio experiences. We included a brief session before the evaluation to inform them about the process of the audio evaluations.

• Sample size: Recent related works like [R1] use 12 users for their perceptual study. We understand your concern and further extend our user study from 15 users to 26 users with an updated result shown below. Our method still demonstrates superior perceptual preference across all aspects.

• Additional evaluations: We also incorporate an objective, reference-based audio similarity metric Deep Perceptual Audio Metric (DPAM) as suggested by reviewer eBxj. This metric can evaluate the audio similarity close to human perception. We evaluate the similarities between the rendered sound (IR convolve with music) to the ground truth. As shown in the table, Versa-ELE reduces DPAM error by 20%, and Versa-SSL further reduces error by 36% compared with Versa-ELE on the AVR method. Versa-ELE also boosts the performance of other methods by 18%. These show our method has stronger perceptual similarity to ground truth. These results are consistent with the user study findings and reinforce the perceptual benefit of our method in the resounding task.

We will revise the manuscript to clarify the study design, IRB waiver and add the new evaluation metric to enhance the robustness of our perceptual analysis.

[R1] Tong, G., Leung, J. C. H., Peng, X., Shi, H., Zheng, L., Wang, S., ... & Chakravarthula, P. (2025). Multimodal Neural Acoustic Fields for Immersive Mixed Reality. IEEE Transactions on Visualization and Computer Graphics.

---

Minor comments:

Typos: Thanks for pointing out the typos and we will fix them in the revised version.

What does L81 mean? It highlights the growing importance of RIR modeling in immersive audio rendering for virtual and augmented reality applications. RIR characterizes how sound propagates through an environment, which is essential for generating spatially and temporally realistic audio cues. These cues greatly enhance user immersion in virtual environments, aligning with the motivation behind our work on neural acoustic fields and the resounding task.

(Continue of the previous discussion, this is the part two of the response)

We do agree that the perfect reciprocity of RIRs can not be satisfied in the real-world. And we aim not to make the strong assumption or claim that reciprocity is naturally applied to any real dataset. To avoid this overclaiming, we will revise our wording throughout the paper to clarify that Versa does not assume strict reciprocity in the data, but rather leverages reciprocity-inspired constraints as an inductive bias in modern neural acoustic field training. Our empirical results also show that with our reciprocity inspired Versa method, each model can get better performance in the acoustic field reconstruction. This shows that injecting these physical properties can make the NN model learn a more generalizable acoustic field. In addition, we will replace phrases like "enforces reciprocity" with "encourages reciprocity-based consistency" or "incorporates reciprocity as a soft constraint.” Finally, as also noted in [R4], the reciprocity of a practical real system needs to be checked to eliminate the doubt. We would also add text to advise that before using Versa on the dataset, one can first check if the reciprocity of RIRs exist in their setting first.

Furthermore, we will add the following clarifying paragraph in our method section and introduction:

"While the principle of reciprocity holds under idealized conditions (e.g., linear, time-invariant media with symmetric reflection), real-world environments and simulated systems may deviate from these assumptions due to many factors like measurement noise, nonideal hardware responses, or complex material properties and so on. Rather than assuming perfect reciprocity in the data, Versa uses reciprocity as a guiding principle to regularize the learning process by encouraging output consistency when emitter and listener positions are exchanged. This approach allows the model to benefit from symmetry in wave propagation, even when perfect reciprocity does not strictly hold."

Further clarification of our motivations at Line 81

We appreciate the reviewer’s historical perspective. We agree that convolution with an impulse response has been a foundational rendering technique for over 50 years, and it remains the physically correct way to simulate how sound propagates through a linear, time-invariant medium. Our intention with this statement was to highlight the renewed interest in the way to render impulse response within the deep learning and neural acoustic rendering topics (NAF and INRAS (NeurIPS’22); AV-NeRF (NeurIPS’23), DiffRIR (CVPR’24), AVR and AV-Cloud (NeurIPS’24), NeRAF (ICLR’25), and more papers on recent ICCV, ICML submissions), especially in the context of spatial audio synthesis for immersive experiences. We will rephrase the sentence to better reflect this nuance:

"Impulse response modeling has seen renewed interest in recent years, particularly in the context of neural acoustic field learning and immersive spatial audio rendering for virtual environments."

Summary with proposed manuscript revisions:

Again we would like to thank you for your engagement to help us improve the scientific rigor of our paper. To address the your key concern, we will make the following changes in the revised manuscript:

• Adjust phrasing throughout to reflect that Versa uses reciprocity as a soft inductive bias.
• Add a subsection discussing the assumptions and limitations of reciprocity in real/simulated datasets.
• Include references to acoustic modeling literature acknowledging material and device nonidealities. We also will add more references to discuss the reciprocity in a practical real-world acoustic system.
• Add the additional experimental results to demonstrate the RIRs similarities of paired one and cross-paired one both in our self-collected real dataset and the simulated dataset.
• Clarify our terminology around RIR modeling’s historical context and our paper’s motivations.

References

[R1] Helmholtz, H. von. "Theorie der Luftschwingungen in Röhren mit offenen Enden." (“Theory of air vibrations in € pipes with open ends”) (1860): 1-72.

[R2, cited in 38 in the manuscript] Rayleigh, John William Strutt Baron. The theory of sound. Vol. 2. Macmillan, 1896.

[R3] Hamilton, M. F., & Blackstock, D. T. (2024). Nonlinear acoustics (p. 447). Springer Nature.

[R4] Ten Wolde, Tjeert. "Reciprocity measurements in acoustical and mechano-acoustical systems. Review of theory and applications." Acta Acustica united with Acustica 96, no. 1 (2010): 1-13.

[R5] Baba, Youssef EI. Room Geometry and Low-Frequency Transfer Function Inference Using Acoustic Transducers. Friedrich-Alexander-Universitaet Erlangen-Nuernberg (Germany), 2022.

[R6] Lu, Ling, Hongling Sun, Ming Wu, Qiaoxi Zhu, and Jun Yang. "A modified electro-acoustical reciprocity method for measuring low-frequency sound source in arbitrary surroundings." Applied Acoustics 116 (2017): 1-8.

(This is the first part of the response)

Dear reviewer, thank you for your thoughtful and detailed follow-up comments and especially your appreciation about our additional experimental results. We are very grateful for your feedback in helping us improve the scientific clarity and rigor of our work. Below, we further address your concerns and propose concrete revisions to the manuscript.

Clarification on reciprocity assumptions and dataset validity

As the reviewer suggests, we do agree that the perfect reciprocity of RIRs recordings can not be satisfied in the real-world. Thus the strict reciprocity of RIRs cannot be assumed or proved in our selected dataset. Instead of proving or assuming perfect reciprocity in the real-world, our method seeks to leverage reciprocity-inspired inductive bias in modern neural acoustic field training. We aim to solve potential misunderstandings and avoid overclaim, providing the detailed reasoning as follows.

Reciprocity of the acoustic system was first proposed by Helmholtz [R1] and expanded by Rayleigh in the theory of sound [R2]. As a theoretical principle, it relies on assumptions such as linearity, time-invariance, and symmetric propagation. As you pointed out, we understand that in real-world recordings, perfect global reciprocity, i.e. impulse response invariance after switching emitter and listener positions, is hard to achieve due to all factors including device non-linearity, non-linear material responses, hardware thermal noise. We will also cite relevant work [R3] to acknowledge the non-ideal behaviors of some common materials under high sound energy, while noting that in the low-intensity, indoor acoustic settings, quasi-linear response is a widely accepted approximation. We also note that, as you suggested, even in a simulated environment, the reciprocity does not strictly hold due to the numerical calculation errors, randomness in the ray-tracing and so on.

In fact, as also presented in our additional real-world RIR collection experiments, the distance between the paired RIRs is not strictly zero, the recording is naturally influenced by the hardware thermal noise and all the non-perfect properties in the real world mentioned before. However, the paired RIRs still have much higher similarities compared with the cross-paired RIRs. To validate the reciprocity on the simulated dataset, we also include an additional experiment on RIRs similarities on the AcoustiX dataset. Results show that with the increased number of rays, the error introduced by the numerical results on the reciprocity gets decreased since there is less randomness in the higher order reflections in the ray tracing procedure. Overall the results indicate that the RIRs are very close(100X) after emitter/listener swapping with 10^6 rays on the simulated dataset. Again the RIRs are not fully identical (metrics are not 0), but through the waveform visualization (which we will incorporate in the revised version) we found that the RIRs are nearly the same. We also adopt this 10^6 rays setting in the simulation settings. Further increasing this number in the simulation would cause very high computational and memory overload.

We also understand your concern about whether perfect reciprocity holds on the impulse response. While the linearity, symmetric propagation and other requirements are needed to satisfy the global reciprocity and the impulse response invariant features, real world collection can inevitably encounter factors like noise and non-linearity, resulting in imperfect impulse response that could “break” the reciprocity. Prior work has demonstrated that reciprocity may not hold perfectly due to speaker imperfections. However, with proper measurement techniques and deviation correction, reciprocity remains valid and reliable in real-world acoustic environments, even in reverberant settings [R6]. Suggested by [R4], even when slight disturbances exist, reciprocal measurement setups still yield data with acceptably small uncertainty. Also the acoustic path reciprocity has facilitated the impulse response measurements by allowing for switching the positions of emitter and listener without affecting the measurement, thereby enabling measurement scenarios where loudspeaker placement is potentially challenging and microphone placement is straight forward [R5]. A similar idea has also been leveraged in our paper to introduce the reciprocity inspired Versa method as a learning strategy.

We thank Reviewer tDBR for the constructive suggestions. The feedback has significantly contributed to the improvement of our paper to make it better accessible to the broader community. We appreciate the reviewer for recognizing that “the combination of physical modeling sophistication with machine learning sophistication is a strength”, and “the improvements over baselines appear substantial”, giving us a positive rating. We address your concerns as follows.

---

Reivew W1: Concern about the scope and accessibility.

We understand the reviewer’s concern about accessibility to the broader NeurIPS community. While the application domain is audio and acoustics, our contribution is fundamentally about machine learning under a structured physical prior. Specifically, we propose reciprocity-grounded Versa as a learning strategy. We show the effectiveness of Versa in many existing work on audio field reconstruction that are published in the NeurIPS community (NAF, INRAS, AVR, and AV-NeRF). We also provide theoretical grounding and empirical evidence for the proposed method for learning the dynamic audio field.

We hope the broader NeurIPS audience finds value in how our method bridges wave propagation and ML training pipelines. We will highlight these broader themes and revise the introduction and discussion to explicitly call out these cross-domain insights.

---

Reivew Q1: Adding pointer in the demo.

Thanks for the advice, we will add extra pointers to highlight the comparisons and point-by-point comparison for the audio after releasing our work on the project webpage.

---

Reivew Q2: Question about the rendered sound, how about the rendering speech?

We appreciate this thoughtful question. Our method fundamentally models impulse responses, which characterizes how sound waves propagate in the environment and heard by human ears. Since the impulse responses are independent of the input dry sound, it can be convolved with any sound source like music, speech and ambient noise. As a result, the resounding capability generalizes naturally across genres and source types.

While we used music in the demos for its perceptual richness and clarity in spatial cues, our method works equally well with speech and other sounds. We will include additional speech-based examples in the supplementary material and project website to illustrate this.

---

Reivew Q3: How does the approach interact with the physical model?

The physical characteristics of surfaces like sound reflections directly influence the room’s impulse response. Integrating explicit material priors, either from semantic scene understanding or estimated material parameters is a promising direction for future work. This could further enable data efficiency by grounding the acoustic model in known physical properties. Given the material’s properties, one can use a ray tracing method to render the impulse response, where the reciprocity property is already considered.

Versa is complementary to these efforts. While material-based models focus on surface-level interactions, our method emphasizes the broader wave propagation paths, enforcing physical consistency across emitter-listener pairs. It leverages reciprocity to capture the cumulative effects of symmetric path interactions along surfaces for better acoustic field reconstruction. Besides, some of our baseline models (AV-NeRF, INRAS, NeRAF) include relevant physics of the environment from vision or 3D inputs. Experiment results show that Versa improves their performance substantially as well.

---

Reivew Q4: Broader take home message.

We thank the reviewer for encouraging us to better articulate the broader message. At a high level, our work demonstrates how physical priors like reciprocity can be converted into effective training strategy, especially when training samples are sparse. Within the audio domain, we show the effectiveness in many works published in the NeurIPS community.

Beyond the audio domain, this principle extends beyond acoustic signal to vision (light transport), radio frequency (RF) signal and so on. Similar ideas can guide neural rendering models to enforce multi-view consistency. For RF signals, electromagnetic reciprocity holds in many indoor environments. This can also be exploited to simulate virtual transmitters/receivers and reduce measurement overhead in RF channel modeling.

Our key takeaway is the potential of incorporating physical principles into the training of neural networks. This approach not only reduces the 'black box' nature of the models but also enhances their generalization capabilities, often requiring less training data.

We thank Reviewer 9Edj for the careful review and constructive suggestions. The constructive feedback has significantly contributed to the improvement of our paper. We are glad that the reviewer found our work “well presented and motivated” and “well formalized and executed”, giving us a positive rating. We address your questions and concerns as follows.

---

Review W1: Concern about the paper's originality on the reciprocity theory.

We appreciate the reviewer’s point. While reciprocity is a fundamental property in the wave propagation theory. To our knowledge, no prior work has explicitly applied emitter/listener swapping as a data augmentation strategy for training neural acoustic field models. Versa-ELE is effective and novel in its systematic application and broad compatibility across models to solve the resounding task. Moreover, our key contribution lies in extending beyond the same gain pattern case with Versa-SSL, which handles directional asymmetries through a physically grounded self-supervision learning. We will clarify this distinction in the revised manuscript.

---

Review W2, Q1: Question on the permutation invariant structure.

Our method is grounded in the physical property of acoustic reciprocity. The impulse responses remain unchanged when emitter and listener positions are swapped, but only under certain conditions, such as omnidirectional gain patterns. This condition is important when evaluating whether architectural permutation invariance is an appropriate solution.

1. Discussion on permutation invariance architecture: Based on the reciprocity principle, it is possible to design permutation-invariant neural networks (e.g., with shared encoders for emitter/listener poses), such architectures enforce symmetry unconditionally, regardless of whether impulse response remains the same with direct swapping. However, emitters and listeners can have different directional gain patterns, which breaks the impulse response invariance (discussed in the sec3.3). Enforcing symmetry in such cases can encourage the model to make incorrect assumptions. This is also the motivation for the development of our proposed Versa-SSL in the case of different emitter/listener gain patterns. Versa-SSL shows improved performance in these cases that simple permutation invariant networks cannot achieve.

2. The advantage of Versa: Versa-ELE explicitly leverages reciprocity when it holds by generating physically valid virtual samples through pose exchange. This strategy is effective, model-agnostic, and compatible with any architecture, which does not require changing the model architecture. Furthermore, Versa-SSL generalizes this idea to asymmetric gain patterns by introducing self-supervised consistency constraints, without assuming symmetry.

3. Additional Experiment results: We provide experimental results (on AcoustiX-Diff with different gain patterns) on a permutation-invariant version of NAF. As shown in the table, permutation-invariant structure can improve the performance on the original NAF model under the sparse emitters settings. As this design treats the emitter/listener exactly the same, it can alleviate the problem of sparse emitters. However, the permutation-invariant design strictly forces the model to output symmetric results and does not consider the unique features of the emitter or listener, i.e. different gain patterns. In contrast, vanilla neural acoustic models that encode emitter/listener features separately are not strictly constrained to this inappropriate symmetry and can benefit from the Versa-ELE for better performance. Finally, AVR with Versa-SSL achieves the best overall performance as it correctly handles the different gain patterns.

---

Review Q2: Confusion on equation 8: $F(P_l, P_e, w_l, w_e, G, G) = F(P_e, P_l, w_e, w_l, G, G)$

G is used as both the emitter gain pattern and the listener gain pattern, thus appearing twice in the equation. In the implementation, the emitter gain pattern $G_e$ is implicitly encoded in the neural acoustic field, and it is not available to be changed. Thus, we can only align the listener pattern with $G_e$, forcing the acoustic volume rendering step of AVR to use the same listener gain pattern as the implicit emitter pattern. As a result, we only alter the microphone pattern (we replace the microphone pattern with $G_e$ during the acoustic volume rendering) in the self-supervision stage.

---

Reivew Q3: Clarification on the L200: As a result, Versa-SSL does not generate new valid training samples with ground-truth impulse response as Versa-ELE.

Sorry for the confusion in this sentence. For Versa-ELE, it is implemented as a data augmentation method to generate valid impulse response training samples. For Versa-SSL, it forces the neural acoustic field to output consistently after exchanging the emitter/listener poses in a self-supervised way.

---

Review Q4: Discussion of the relevant paper.

We thank the reviewer for pointing out the work by Ribeiro et al. (ICASSP 2023). While that paper discusses reciprocity in the context of kernel interpolation and physics-informed modeling, our approach differs in both scope and methodology. No prior work has leveraged reciprocity as a systematic training strategy for the neural acoustic field. Our contribution lies in using reciprocity to improve model generalization in sparse emitter settings, and in extending this idea to asymmetric scenarios through Versa-SSL. We will update the related work section to better clarify this distinction.

---

Typos/Suggestions.

Thank you for your detailed suggestions on the writing and text. We will incorporate them for the revised version after rebuttal.

We thank Reviewer eBxj for the thoughtful review. The constructive feedback has significantly contributed to the improvement of our paper. We appreciate the reviewer for recognizing that our work deals with a problem that prior methods “are unable to fully address at the onset”, “can be combined with existing approaches”, and “show improvement in multiple metrics”. We address your questions and concerns one by one as follows.

---

Review W1 part1. Discussion on the related paper BEE.

We thank the reviewer for pointing out this relevant line of work on dynamic acoustic field reconstruction. We agree that BEE[R1] is an important contribution in the broader space of learned acoustic field modeling. While we did not include this work in our original related work section, we will add it in the revised manuscript and clarify how our setting differs as follows.

1. Training Setup: BEE trains on a large-scale, multi-scene corpus with dense emitter-listener pair sampling across a variety of environments. It also leverages audio-visual correspondence as a strong learning signal. After training on the large-scale dataset, it can leverage the sparse listener samples in the novel scene to reconstruct the acoustic field. In contrast, our method focuses on per-scene optimization, only trained on a sparse set of emitter positions, and does not consider cross-scene generalization. Our goal is to reconstruct a physically consistent acoustic field within a single environment, given limited emitter setups.

2. Task definition and output: BEE targets novel scene generalization, which involves predicting audio waveforms or spectrograms for novel listener positions in the environment given sparse receiver samples. In contrast, our work tackles the resounding task in the same acoustic scene, which focuses on estimating the impulse response at an arbitrary emitter/listener location within a scene. This allows for more flexible downstream rendering.

We will revise the related work section to include and discuss our differentiation from BEE and similar approaches.

---

Review W1 part2: Discussion on the related paper AVCloud, NeRAF, INRAS++

Thank you also for pointing out recent advances in acoustic field modeling. We now include two other open-source papers, including AV-Cloud (R3) and NeRAF (R4), as additional baselines in our experiments. Both the AV-Cloud and NeRAF leverage 3D visual information for acoustic prediction.

We experimented with the AcoustiX-Same dataset and applied our Versa-ELE method on top of AV-Cloud and NeRAF. We find that Versa-ELE improves both of their performance by a large amount, further supporting the generality of our approach to various acoustic field approaches.

The INRAS in our baseline model is already implemented as INRAS++ (R2), extending the 3d grid sampling and extra loss components. The other baseline NAF and AVNeRF implementations also follow the RAF paper (R2).

We will cite all of these works (R1–R4) and clarify the baseline settings in the revised version, and expand our related work section to better reflect recent progress in the field.

---

Review W2 part1: Evaluation dataset, “Not applied to the same dataset as the baselines, like SoundSpaces”.

Thank you for raising the dataset issue. For the SoundSpaces dataset, we also noted some time-of-flight errors on the simulated impulse response as prior paper (AVR), which makes it not ideal to test on some models like DiffRIR and AVR. As a result, we use other simulation engines (AcoustiX and GWA) and datasets, plus real datasets (MeshRIR and RAF), with diverse settings to evaluate the model’s performance.

---

Review W2 part2: Concern on the number of evaluation scenes.

As pointed out by the Reviewer sQL3 in the strengths section, we used various simulated datasets, including ray-based, wave and ray hybrid simulation datasets, and also real datasets to test our method. Our evaluation includes various simulated scenes (3 for AcoustiX, 3 for GWA) and real scenes (MeshRIR, 2 RAF scenes), with a total of 9 scenes. Our experiment also involves various setups for gain patterns, resulting in more than 15 different scene+gain pattern combinations for evaluations. We also incorporate a diverse number of emitters/listeners in each scene to evaluate the robustness of our method. Through these diverse evaluations, we have verified the effectiveness of our Versa framework in various models that purely rely on neural acoustic feature (NAF), scene geometries (INRAS), acoustic propagation phenomenon (AVR), visual-acoustic correspondence (AV-NeRF, AVCloud, NeRAF).

---

Review Q1: Confusion on the self-supervision loss.

We do not involve negative pairs for self-supervised learning as common contrastive loss. For our self-supervision loss, let $h_1 = F(P_l, P_e, w_l, w_e, G, G)$ and $h_2 = F(P_e, P_l, w_e, w_l, G, G)$. The SSL loss $L_{a-ssl}$ only forces $h_1$ and $h_2$ to be identical. For the simplest version, we can directly optimize on the L1 loss between the waveform $h_1$ and $h_2$, i.e., $L_{a-ssl} = |h_1 - h_2|$. In our implementation, we follow a similar strategy to calculate the L1 distance between $h_1$ and $h_2$ in time and spectrogram domain as in the previous framework.

---

Review Q2: Include DPAM metric for evaluations.

Thank you for your suggestion! To evaluate perceptual quality, we computed the Deep Perceptual Audio Metric (DPAM) on the audio clips rendered for user study and demo. These clips were generated by convolving predicted RIRs with dry sound and comparing them to ground-truth renderings.

Versa-ELE reduces DPAM error by 20%, and Versa-SSL further reduces error by 36% compared with Versa-ELE on the AVR method. Versa-ELE also boosts the performance of other methods by 18%. These show our method can improve the perceptual similarity to ground truth. These results are consistent with the user study findings and reinforce the perceptual benefit of our method in the resounding task. We will include this new metric and its analysis in the revised version.

---

Review Q3: Evaluation of Versa on BEE.

We thank the reviewer for the follow-up question regarding the comparison to BEE and the potential for combining Versa-ELE with that framework. We address this question as follows:

1. Different Settings and Objectives: Echoing our response to W1-part1, BEE and our work target distinct problem settings. BEE is designed for cross-scene generalization: it is trained on a large-scale, multi-scene dataset with dense emitter-listener pair sampling, and it leverages audio-visual correspondence to model the acoustic field. After large-scale training, it can be adapted to novel scenes using sparse listener observations. Versa is specifically designed to improve generalization of neural acoustic fields within a single scene, leveraging reciprocity as a physics prior for improved modeling of wave propagation. We also contain only a few emitter locations training dataset, instead of densely sampled emitter-listener pairs in lots of scenes as in BEE.

2. Evaluation on other related models: While we do not compare directly to BEE, we evaluate Versa on other recent open-source neural acoustic field methods that also use audio-visual correspondence, including AVCloud and NeRAF/AV-NeRF. These methods operate within a single scene and use visual priors for field reconstruction. Our results demonstrate that Versa is compatible with such architectures and consistently improves performance.

Dear reviewer,

Thank you for your positive feedback and for recognizing our method's reciprocity-inspired learning strategy. We will update the manuscript according to all your suggestions. We hope our revisions and the demonstrated results further justify your support for our paper's acceptance.

Dear reviewer,

We would like to thank you for the time and constructive feedback during the review process. We have addressed each of your further comments in detail and proposed corresponding revisions to the manuscript. If you have any further questions, clarifications, or suggestions, we would be happy to address them before the discussion period concludes. Thank you again for your engagement. Your suggestions are highly valuable and would greatly help us improve our work.