acceptodds
Under review as a conference paper at ICLR 2027

Hearing Anywhere from Anything: Geometry-Conditioned Acoustic Field Synthesis from Unlocalized Blind Recordings

Abstract

Reconstructing a room’s acoustic field is essential for rendering spatial audio at previously unobserved source locations. Yet existing methods typically require either scene-specific optimization or carefully measured room impulse responses (RIRs) with calibrated positions. We observe that ordinary recordings in everyday life, such as speech and music, are also shaped by the room and therefore contain acoustic evidence. Exploiting this evidence is challenging because both the dry source signals and the positions of the recorded sources are unknown. Our key insight is to formulate it as a physics-anchored conditional completion problem. Unlocalized recordings reveal room-level decay and reverberation, whereas target geometry anchors direct-path timing and specifies the source–receiver path of interest. Building on this insight, we introduce PHILAC, a two-stage framework that first predicts a query-conditioned energy-decay profile and then generates the target RIR through masked discrete diffusion of acoustic codec tokens. On unseen synthetic rooms, PHILAC achieves a mean EDC error of 2.660 dB, improving over the strongest baseline by 19.5%. On real rooms, we outperform the strongest baseline by 45.7% in mean EDC error without scene-specific adaptation. These results show that everyday recordings can provide acoustic context for synthesizing RIRs at new locations without calibrated acoustic measurements.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Acoustic Field Reconstruction, Generative Models, Spatial Audio

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_TGJ2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PHILAC for synthesizing room impulse responses (RIRs) at unobserved source positions. The input is a set of ordinary, unlocalized monaural recordings from the same fixed receiver, a receiver-centered depth panorama, and a target source coordinate. The dry signals and source positions of the context recordings are unknown. PHILAC adopts two-stage generation: a query-conditioned energy-decay planner predicts the target decay profile, then a masked discrete-diffusion renderer fills in the RIR waveform as acoustic codec tokens. Experiments report improvements on six acoustic metrics over adapted baselines on unseen synthetic rooms (AcousticRooms dataset) and zero-shot transfer to four measured rooms (HAA dataset).

Critical strengths

S1. The problem setting is meaningful. Replacing calibrated context RIRs with ordinary, unlocalized recordings substantially reduces the acoustic measurement burden at inference time. The combination of blind acoustic context and known query geometry is practically relevant to fixed-viewpoint audio editing and AR/VR rendering. Using depth as the geometric input is also well motivated: room geometry is often far cheaper to obtain than impulse-response measurements.

S2. The paper has well-motivated architectural decomposition. Separating decay estimation from waveform generation provides an interpretable intermediate representation that can be independently inspected and supervised. Permutation-invariant context aggregation, monotone decay planning, and geometry-derived temporal alignment are reasonable inductive biases. The formulation appropriately acknowledges that the available observations do not uniquely determine fine-scale waveform details, giving a principled motivation for a generative model.

S3. Consistent empirical improvements within the reported protocol. PHILAC improves all six reported metrics on unseen synthetic rooms, with a 19.5% reduction in mean EDC error over the strongest evaluated baseline. Increasing the number of unlocalized recordings from K=1 to K=8 consistently improves performance across all metrics (Table 3), directly supporting the claim that unlocalized recordings accumulate useful room evidence. The real-room improvement persists under equal room weighting (Table 12), and the source-domain and noise experiments (Tables 8, 13) provide useful evidence about generalization boundaries and failure conditions.

Critical weaknesses

W1. The experiments do not sufficiently test whether the model reconstructs a spatially consistent acoustic field.

The paper shows that pointwise acoustic statistics improve, but does not establish that the model recovers spatial structure, e.g., occlusion-aware, position-to-position variation that distinguishes a true acoustic field from a room-average response scaled by distance. Three specific gaps: (a) Could distance alone explain the spatial trends? A simpler model that learns a single room reverberation tail and adjusts direct-to-reverberant ratio by source–receiver distance would also produce DRR maps that look higher near the receiver and lower far away. To rule this out, the evaluation needs queries at similar distances but different directions or visibility, compared against a distance-only baseline. The geometry ablation (Table 5, EDC 2.660 --> 3.998) confirms the branch helps, but does not tell us whether the gain comes from room-scale cues or from fine-grained wall/occlusion awareness. Whether this ablation was retrained is also unclear. (b) What happens when context recordings come from mixed propagation conditions? Real recordings are not all alike: some may have strong direct sound, others may be fully occluded. The model aggregates them with permutation-invariant attention, but this does not guarantee it separates room-level properties from path-specific effects. A small controlled test would show whether the model extracts transferable room evidence or is biased by the input distribution, specifically (Line-Of-Sight(LOS)-only contexts, Non-Line-Of-Sight(NLOS)-only contexts, and a mix, each evaluated on both Line-of-Sight (LOS) and Non-Line-Of-Sight (NLOS) targets. The Real Acoustic Fields 1 dataset is well suited here: it has acoustically distinct connected spaces (large soft-walled room vs. small concrete room), occluded source positions, and measurable T60/DRR differences across regions. (c) Are neighboring predictions consistent along a trajectory? Each query is generated independently (acknowledged in Appendix E). Shared room tokens provide a common condition, but independent sampling of reflections and reverberation detail can still produce discontinuities from one position to the next. A dense-trajectory evaluation would directly test this, e.g. checking whether predicted changes in arrival time, early energy, and gain match the ground truth's changes. The concern is especially relevant to the paper's motivation of rendering moving sources.

W2. The paper does not show that audio rendered with its predicted RIRs is perceptually closer to the real target.

The reported metrics capture decay and energy-ratio statistics but do not directly assess what a listener hears. Two RIRs with identical EDC can sound noticeably different. This gap is especially consequential because the model uses stochastic generation to fill in waveform details the conditioning does not determine, yet Table 5 shows deterministic decoding achieves comparable or better scores on most metrics. Without perceptual evidence, the practical benefit of the stochastic sampling remains unclear.

W3. The evaluation metrics can hide important reconstruction errors.

The six reported metrics primarily measure decay characteristics and energy ratios. All six are invariant to global gain scaling: multiplying an RIR by any positive constant leaves normalized EDC, T60, EDT, C50, and DRR unchanged. A prediction can therefore achieve excellent scores across all metrics while rendering the source at the incorrect loudness level. This is a concrete gap because the model explicitly predicts a gain token, yet no metric validates it.

[1] Chen et al. "Real Acoustic Fields: An Audio-Visual Room Acoustics Dataset and Benchmark", CVPR 2024.

Questions for authors

Q1. Spatial validity beyond distance (W1). Can the authors evaluate queries at similar source–receiver distances but occlusion conditions and compare against a room-and-distance-only baseline? Please also clarify whether the geometry ablation (Table 5) is retrained and which coordinate/delay inputs it retains.

Q2. Heterogeneous context paths (W1). Could the authors provide a small controlled evaluation, e.g., using Real Acoustic Field Dataset as a real-world testing dataset, with input condition LOS-only, NLOS-only, and mixed contexts, reporting results separately for LOS and NLOS targets?

Q3. Perceptual quality (W2). Do the acoustic-metric improvements translate into perceptually more faithful rendered audio? A blind listening comparison using the same dry signals convolved with measured and predicted RIRs would directly address this., i.e., comparing PHILAC, the strongest baseline, and the deterministic variant on both LOS and NLOS targets. At minimum, are audio examples available?

Q4. Gain validation (W3). Since the model explicitly predicts a gain token, why is no loudness or gain-error metric reported?

4 comments

Response to Reviewer TGJ2 I (TBC)

We thank Reviewer TGJ2 for the thoughtful and constructive review, and for recognizing the motivation, architectural decomposition, and empirical consistency of Philac. Below we address each weakness (W1–W3) and question (Q1–Q4) point by point.


W1(a) & Q1: Spatial structure beyond source–receiver distance & geometry ablation clarification

To rule out the hypothesis that Philac merely scales a room-average response by source–receiver distance, we construct a controlled matched-distance occlusion test. Each pair shares the same receiver and nearly identical source–receiver distance (Δd≤0.2\Delta d \le 0.2 m, angular separation ≥60∘\ge 60^\circ), but has opposite ray-mesh visibility (one LOS and one NLOS source). We evaluate the signed parameter change from LOS to NLOS and its pairwise Mean Absolute Error (MAE) against ground truth:

LOS →\rightarrow NLOS contrastMeasuredPhilacFLACEigeNetBlind T60T_{60} + FDN
Mean Δ\DeltaDRR (dB)−3.276-3.276−4.221-4.221−2.081-2.081−1.550-1.5500.0000.000
Δ\DeltaDRR MAE (dB)—2.8062.9973.6574.118
Mean ΔC50\Delta C_{50} (dB)−3.546-3.546−3.555-3.555−2.037-2.037−1.781-1.781−0.009-0.009
ΔC50\Delta C_{50} MAE (dB)—2.3042.8192.9133.790
Mean Δ\DeltaEDT (s)+0.146+0.146+0.139+0.139+0.081+0.081+0.059+0.059−0.001-0.001
Δ\DeltaEDT MAE (s)—0.1050.1150.1220.158

From the results, we observe that Philac achieves the lowest pairwise difference MAE across all three metrics, closely matching the ground-truth mean shifts. Against the room-and-distance baseline (Blind T60T_{60} + FDN), Philac reduces room-weighted difference MAE by 1.432 dB for DRR and 1.664 dB for C50C_{50}.

Clarification on Table 5 geometry ablation: In submitted Table 5, w/o visual/relative geometry masks out the panorama and relative-surface features at inference from the trained full checkpoint while retaining the geometric direct-arrival delay τq=∥q∥/c\tau_q = \|q\|/c; it is not a separately retrained geometry-free model.


W1(b) & Q2: Heterogeneous context paths

We evaluate how context path composition affects prediction by holding the target queries and context budget (K=2K=2) fixed while varying whether the context recordings come from LOS-only, NLOS-only, or mixed (1 LOS + 1 NLOS) positions. We test both unseen synthetic rooms across AcousticRooms and measured real-room RIRs from HAA:

Dataset & target stratumLOS-only context EDC (dB)NLOS-only context EDC (dB)Mixed context EDC (dB)
AcousticRooms, LOS targets3.6023.7703.701
AcousticRooms, NLOS targets3.8273.6593.913
HAA, LOS targets2.9552.7702.800
HAA, NLOS targets1.9321.5531.712

Context composition measurably shifts predictions rather than collapsing to a single context-agnostic prior. On synthetic rooms, matching the context visibility to the target stratum yields the lowest EDC error at 3.602 dB for LOS targets and 3.659 dB for NLOS targets. On the measured HAA rooms, NLOS-only context achieves the lowest EDC error for both target strata because occluded recordings are dominated by the diffuse reverberation tail that constrains room-level EDC.


W1(c): Consistency along dense trajectories

To test spatial continuity under independent per-query generation, we evaluate fixed-receiver, fixed-context (K=8K=8) predictions along three measured HAA trajectories:

Adjacent-Change MAEPhilacFLAC
First onset (ms)0.3661.510
Onset-aligned early energy E50E_{50} (dB)0.9421.299
Peak gain (dB)1.5112.290

We also evaluate 22 mesh-checked grid trajectories on AcousticRooms, where Philac outperforms FLAC on adjacent T60T_{60} change at 0.065 s versus 0.095 s and peak-gain change at 2.158 dB versus 2.599 dB. At the same time, we agree that independent per-query sampling does not by itself guarantee smooth waveform-level continuity for moving-source rendering. We have reported these measured trajectory-change statistics in Appendix C.3 (Table 10) and framed autoregressive or joint multi-query generation as a key next step in Appendix D of the revised manuscript.

Response to Reviewer TGJ2 II

W2 & Q3(a): Perceptual quality

To directly evaluate perceptual quality, we conducted a blinded MUSHRA listening study (ITU-R BS.1534) with 12 participants across 12 trials spanning Speech, Music, and RIR stimuli, comparing Philac, FLAC, EigeNet, and the Hidden Reference on a 0–100 perceptual similarity scale. All 12 participants passed the hidden-reference screening check (rating the Hidden Reference ≥90\ge 90 on every trial):

ConditionMean (95% CI)StdRangeQuality Tier
Hidden Reference95.40±0.4095.40 \pm 0.402.4590–100Excellent
Philac (Ours)85.42±0.9385.42 \pm 0.935.6874–96Excellent
FLAC76.40±1.8776.40 \pm 1.8711.4252–91Good
EigeNet56.83±2.6056.83 \pm 2.6015.9425–85Fair

Philac reaches the Excellent perceptual quality tier and outperforms the strongest baseline FLAC by +9.02 points with substantially lower variance across listeners and stimuli. We have included the full MUSHRA results and evaluation protocol in Appendix C.4 (Table 11 and Figure 9), and interactive audio examples are available in the paper.


W2 & Q3(b): Role of Stochastic Sampling

As also noted by FLAC (Brunetto, 2026), sparse acoustic observations leave fine late reflections underdetermined. Table 5 and Table 2 reflect a clear operating tradeoff between deterministic and stochastic decoding:

  • Early geometry vs. late decay in Table 5: Deterministic decoding slightly favors early-energy ratios anchored by query geometry, reaching 1.857 dB C50C_{50} and 1.733 dB DRR versus 2.031 dB and 1.831 dB. Stochastic sampling improves full and late decay, lowering EDC from 2.771 to 2.660 dB and Late EDC from 3.330 to 3.216 dB by avoiding mode averaging in the reverberant tail.
  • Test-time selection in Table 2: A stochastic conditional distribution enables Best-of-NN blind selection with N=8N=8, which outperforms deterministic decoding across both early and late metrics: 2.543 dB EDC, 3.095 dB Late EDC, 17.56% T60T_{60}, 1.764 dB C50C_{50}, and 0.0848 s EDT.

W3 & Q4: Validation of absolute gain and loudness

We agree that the six normalized metrics in the main tables are invariant to global scaling and do not validate the predicted gain token zqgz_q^g. We therefore evaluated pre-normalization peak RIR levels:

MethodPeak MAE (dB)Room-macro MAE (dB)Within 3 dB (%)
Philac2.1291.75577.4%
EigeNet2.7962.72665.6%
FLAC3.4042.50260.9%

Philac achieves the lowest peak-level error and places 77.4% of predictions within 3 dB of ground truth. We have included this pre-normalization peak gain validation in Appendix C.13 (Table 19) of the revised manuscript.

Dear Reviewer TGJ2,

Thank you again for your thoughtful and constructive review, and for recognizing the meaningful problem setting, two-stage architectural decomposition, and empirical consistency of PHILAC.

As the discussion deadline approaches, we wanted to gently follow up to see if our response and revised manuscript have addressed your concerns. Specifically, we have added the matched-distance LOS-to-NLOS occlusion test (Appendix C.3, Table 8), heterogeneous context visibility and dense trajectory evaluations (Appendix C.3, Tables 9–10), the blinded ITU-R BS.1534 MUSHRA listening study with interactive audio examples (Appendix C.4, Table 11), and pre-normalization peak gain validation (Appendix C.13, Table 19).

We would welcome any further feedback and remain available to address any remaining questions during the discussion period. Thank you for your time and effort!

Best regards, The Authors

Dear Reviewer TGJ2, we kindly reach out before the discussion window closes to see if our new spatial field analyses and MUSHRA listening evaluation have addressed your concerns, and we welcome any additional feedback.

Reviewer_52QVOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper tackles few-shot RIR estimation with extra relaxations on audio references. The references are now ordinary signals convolved with RIRs instead of RIRs themselves, and no (relative) source positions are given. An egocentric panoramic depth from a query receiver is also given, following FLAC.

This setting is not the most general or practically applicable one; it still assumes a single static, omnidirectional source in each recording and does not allow inference at unseen receiver positions (which I see as a missed opportunity). Nevertheless, the current setting should be sufficiently challenging to warrant a standalone investigation.

The authors then proceed to present quite a complex system that combines a reference encoder, query (depth) encoder, EDC predictor, and a discrete diffusion model that operates on a RQVAE-based RIR codec. Results are reported on the synthetic AcousticRooms dataset (Treble wave+path tracing simulations) and the HAA dataset (real-world measurements).

Critical strengths

The proposed work relaxes the FLAC setup in two ways, closing the gap between the lab settings and actual potential application scenarios.

Critical weaknesses

After the problem formulation, the paper immediately jumps into building a complex system and reporting the benchmark numbers. As a result, the current manuscript misses several research questions that I find more interesting. Why and how does the relaxation make the task more difficult? What kind of specific acoustic details do we miss? In which scenario does the task become more difficult? When and why does the baseline struggle in particular? Without these investigations, it is not clear how each part of the proposed model motivates itself over its baseline counterparts.

I believe the baseline comparisons are not fair and should not be the main focus of the paper. It puzzles me how Speech2RIR and Lluis's (2025) Blind SRIR model can be compared with the proposed method. From my understanding, they are not designed for the task tackled here at all; they are not few-shot (K>1) estimators and do not take the panoramic depth as input. FLAC, EigeNet, and xRIR might be a better match, but they still aren't designed to capture room acoustic features effectively from reverberant recordings; they are designed for RIRs.

Questions for authors

I left key questions in the weaknesses section. Below are the remaining questions and additional comments.

Could the authors please elaborate on how the baseline systems were adapted to the presented setup? The appendix seems a bit brief.

I find that at least the following ablation studies are needed. Please correct me if I am wrong.

(1) w/ RIRs & source positions (2) w/o RIRs but with source positions (3) w/ RIRs but without source positions (4) w/o RIRs & source positions (i.e., current setting)

I would like to see a small set of qualitative examples demonstrating the nature of the task and the behavior of the proposed model (plus baseline models). For example, we can pick a challenging scene with many occlusions (e.g., one of the Apartment scenes in AcousticRooms), draw room acoustic parameter and pressure distribution plots, and discuss what is happening there. I find Figures 7 and 9 to be a great starting point.

It seems like the acoustic planner predicts full-band EDC. Did the authors try frequency-dependent ones, e.g., octave-band decays or energy decay relief (EDR)?

Why not enforce (e.g., renormalize) the generated RIR to match the EDC, instead of the best-of-N inference?

While the paper is generally clear, are all math symbols defined in the text (e.g., R_\theta)? Personally, I found those equations redundant and, at times, confusing.

Appendix A can be merged with Appendix B. Appendix C (metrics) seems to be redundant.

Appendix D.6: It seems like more refinement steps do not improve performance. Why? Is this an intended outcome or explainable behavior? Is the decoder truly "generative"?

Appendix D.7: Did the authors retrain the baseline codes for RIRs?

Appendix D.8: Why do we need both query-weighted and room-macro results? What extra insights do we gain here?

7 comments

Response to Reviewer 52QV I (TBC)

We thank the reviewer for the detailed and candid review. Below we address each weakness (W1–W2) and question (Q1–Q10) point by point.


W1 & Q2: Why and how does the relaxation make the task harder?

We agree that these questions should precede and motivate the model design. Therefore, we conducted two diagnostic studies to answer them directly.

1. Controlled breakdown of the two relaxations

We aligned native clean checkpoints and trained strict-blind checkpoints on the same data:

MethodContext observationContext poseEDC (dB)Late EDC (dB)T60T_{60} rel. (%)C50C_{50} (dB)EDT (s)DRR (dB)
FLACClean RIRKnown1.5581.91511.361.1380.05551.387
FLACClean RIRHidden1.5711.93911.391.1360.05731.416
FLACBlind recordingHidden3.3054.05521.982.5080.10432.323
EigeNetClean RIRKnown1.3931.6719.291.2960.04721.840
EigeNetClean RIRHidden1.8402.21111.691.5790.06102.058
EigeNetBlind recordingHidden3.3754.15221.922.7210.12562.328
xRIRClean RIRKnown1.6091.8868.071.3390.05871.931
xRIRClean RIRHidden3.1063.39710.284.1620.13894.703
xRIRBlind recordingHidden6.9228.14942.226.0560.27976.499
OursBlind recordingHidden2.6603.21618.392.0310.09571.831

From these results, we can observe that:

  • xRIR struggles primarily when context positions are removed: Its EDC error rises from 1.609 to 3.106 dB (and C50C_{50}/DRR errors jump from 1.339/1.931 dB to 4.162/4.703 dB) when clean context positions are hidden, because its architecture relies on coordinate-based temporal alignment of reference paths.
  • FLAC and EigeNet struggle primarily when clean RIRs are replaced by blind recordings: Removing only context pose changes FLAC EDC from 1.558 to 1.571 dB and EigeNet from 1.393 to 1.840 dB, whereas switching to blind recordings raises their errors to 3.305 and 3.375 dB. Furthermore, when dry signals lack speech pauses (submitted Table 8), their errors rise sharply on broadband noise (12.751 and 9.914 dB) and music (5.962 and 6.560 dB), because source dynamics mask the room decay.

Why this makes our problem fundamentally harder: Standard few-shot RIR methods receive direct impulse envelopes, unmasked decay tails, and calibrated spatial paths to interpolate from. Removing both clean RIRs and context poses simultaneously strips away impulse observability (i.e., entangling reverberation with unknown, non-stationary dry signals and removing absolute gain references) and spatial grounding (i.e., preventing geometric alignment or interpolation across context paths). Consequently, even after full retraining on the same data, existing few-shot RIR models suffer a 2.1×2.1\times–4.3×4.3\times surge in EDC error (and up to 12.8 dB on continuous excitations). Solving this requires disentangling shared room acoustics from arbitrary dry content and re-anchoring them to an unobserved target path, a fundamentally harder inverse problem than interpolating calibrated RIRs.

2. Stratification by distance and occlusion

We stratify the queries (K=8K=8) by target source–receiver distance and Line-of-Sight (LOS) vs. Non-Line-of-Sight (NLOS) propagation:

Target distancePhilac EDC (dB)FLAC EDC (dB)NLOS-only Philac / FLAC EDC (dB)
<2<2 m2.7663.6005.085 / 4.503
2–42\text{--}4 m2.7763.3534.066 / 4.115
4–84\text{--}8 m2.6993.2233.616 / 3.640
≥8\ge 8 m2.3012.8092.217 / 3.073 (182 queries)
  • Impact of the missing input cues on target acoustics: While room-level late decay (T60T_{60}) remains recoverable from unlocalized recordings (especially at ≥8\ge 8 m where reverberation dominates), the lack of path-specific early-reflection and gain cues in blind inputs makes early-to-reverberant energy contrast (DRR and C50C_{50}) around local occlusions the hardest property to infer (detailed further in Q3).
  • Hardest scenario: Although Philac achieves lower query-mean EDC than FLAC in every full distance bin, the short-range (<2<2 m) NLOS subset is harder for Philac (5.085 vs. 4.503 dB across 34 queries), showing that aggregate gains do not cover every propagation regime. (Note also that the 2–4 m room-bootstrap interval crosses zero and the ≥8\ge 8 m bin spans only 3 rooms, so distance and room composition covary.)

Response to Reviewer 52QV II (TBC)

3. How these findings motivate each part of Philac over its baseline counterparts

  • Blind Context Encoder EctxE_{\text{ctx}} vs. baseline RIR and pose encoders: Unlike xRIR's pose-dependent path alignment or FLAC and EigeNet's RIR waveform encoders, EctxE_{\text{ctx}} uses a BERP-inspired Conformer and pose-free SetAttn with an auxiliary room-parameter head to isolate invariant room cues from arbitrary excitations. This keeps broadband and music EDC errors at 2.899 and 3.129 dB, compared to 12.751 and 9.914 dB on broadband noise and 5.962 and 6.560 dB on music for FLAC and EigeNet in Table 8.
  • Query-Relative Geometry EgeoE_{\text{geo}} and Factored Codec DϕD_\phi vs. generic representations: Because context poses are lost, EgeoE_{\text{geo}} anchors visible surfaces directly to the target source qq via the query-relative surface field MqM_q, and removing this geometry raises EDC error to 3.998 dB in Table 5. Meanwhile, canonical time alignment τq=∥q∥/c\tau_q=\|q\|/c and factored gain gqg_q let the codec model only normalized RIR shape, outperforming DAC at 4.101 dB and EnCodec at 3.287 dB in Table 11.
  • Acoustic Planner PψP_\psi with FiLM and Discrete Diffusion RθR_\theta vs. single-stage or deterministic predictors: Unlike deterministic regression in EigeNet and xRIR or single-stage flow matching from noisy blind features in FLAC, PψP_\psi with FiLM explicitly predicts a monotone query-level decay curve to anchor global energy decay first, cutting memory-less EDC error from 4.316 to 2.789 dB in Table 6. This leaves RθR_\theta to resolve residual one-to-many fine reflection structure, outperforming deterministic decoding at 2.660 versus 2.771 dB in Table 5.

We have incorporated this diagnostic breakdown (Table 2), failure-mode analysis, and motivation-to-design mapping immediately after Section 3.1 in the revised manuscript.


W2 & Q1: Fairness and scope of baseline comparisons

We clarify two points regarding fairness and baseline scope:

  1. No existing baseline matches our exact setting, and adapted baselines are retrained with identical inputs: Because prior work addresses either few-shot synthesis from calibrated clean RIRs or blind deconvolution without scene geometry, no existing method natively shares our task formulation. To build the fairest possible comparison against the closest few-shot models, namely FLAC, EigeNet, and xRIR, we did not evaluate their pretrained clean-RIR checkpoints directly. Instead, we adapted and retrained them under the exact same room split and input availability as Philac through three steps:
    • Context input replacement: We replace the K=8K=8 clean context RIRs with the exact same K=8K=8 blind recordings, namely 3-second reverberant audio clips from the same room.
    • Withholding context positions: Because context source coordinates pkp_k are unknown in our task, we withhold pkp_k while retaining the known target query coordinate qq and the receiver-centered depth panorama GG. Specifically, context coordinate embeddings or path-alignment branches in xRIR, FLAC, and EigeNet were replaced with zero/learned null embeddings while keeping their target-query branches intact.
    • Retraining from scratch: Each baseline preserves its original network architecture, target-query conditioning, and training objective, and was retrained from scratch on the common room-disjoint training split.
  2. Architectural isolation: Because all geometry-aware few-shot models receive identical training splits and identical inputs, their comparison is fair in terms of data and input access. The higher errors of the baselines isolate the architectural difficulty of capturing room acoustics from reverberant recordings rather than clean impulses.
  3. Role of Speech2RIR and Blind SRIR: We agree that Speech2RIR, which performs single-instance blind deconvolution of the observed recording, and Blind SRIR, which uses tetrahedral inputs without scene geometry, are not designed for K>1K>1 geometry-conditioned synthesis. They were included only as cross-task reference points to verify empirically that single-path blind estimation and geometry-free extrapolation cannot solve this task, rather than as direct apples-to-apples competitors.

Response to Reviewer 52QV III (TBC)

Q3: Challenging scene with many occlusions

Following your suggestion, we select a controlled occlusion pair in Apartments_idx_50, comparing an unoccluded Line-of-Sight source against an occluded Non-Line-of-Sight source at equal Euclidean distance from the receiver and a 126.87∘126.87^\circ angular separation:

Acoustic parameter (LOS →\rightarrow NLOS)Measured RIRPhilacFLACEigeNet
DRR (dB)−4.500→−12.693-4.500 \rightarrow -12.693 (Δ=−8.19\Delta = -8.19)−4.472→−10.967-4.472 \rightarrow -10.967 (Δ=−6.50\Delta = -6.50)−4.595→−8.560-4.595 \rightarrow -8.560 (Δ=−3.97\Delta = -3.97)−6.350→−8.000-6.350 \rightarrow -8.000 (Δ=−1.65\Delta = -1.65)
C50C_{50} (dB)5.257→2.9295.257 \rightarrow 2.929 (Δ=−2.33\Delta = -2.33)5.745→3.1175.745 \rightarrow 3.117 (Δ=−2.63\Delta = -2.63)5.706→3.2925.706 \rightarrow 3.292 (Δ=−2.41\Delta = -2.41)7.481→4.9387.481 \rightarrow 4.938 (Δ=−2.54\Delta = -2.54)
  • Occlusion contrast (∣Δ∣|\Delta| error): Ground-truth occlusion causes a sharp 8.19 dB drop in DRR and 2.33 dB drop in C50C_{50}. Philac captures the strong DRR drop with a ∣ΔDRR∣|\Delta\text{DRR}| error of 1.70 dB, compared to 4.23 dB for FLAC and 6.54 dB for EigeNet, and achieves a ∣ΔC50∣|\Delta C_{50}| error of 0.30 dB (close to FLAC's 0.09 dB).
  • Absolute endpoint accuracy & room-wide consistency: Across all four LOS/NLOS endpoints (DRR and C50C_{50}), Philac's summed absolute bias is 2.43 dB, compared to 5.04 dB for FLAC and 10.78 dB for EigeNet. Moreover, across all 250 queries in this Apartment scene, Philac achieves the lowest room-wide mean error on both EDC (2.561 vs. 3.754 / 3.074 dB) and C50C_{50} (1.350 vs. 1.389 / 1.441 dB).
  • What detail remains hardest: In the heavily occluded NLOS state, all methods slightly underestimate the full 8.19 dB direct-energy drop, highlighting that recovering extreme direct-to-reverberant attenuation behind partitions is the primary missing detail. We have included this Apartment case study and its spatial distribution plots in Appendix C.3 (Figure 8, alongside Figure 7) of the revised manuscript.

Q4: Frequency-dependent planner, such as octave-band decays or EDR?

To test this directly, we trained a six-band octave planner pilot at 125, 250, 500, 1000, 2000, and 4000 Hz with the encoder and renderer kept frozen, and used separate broadband and octave-band heads to supply targets for output projection across all 5,233 test queries over 3 seeds. Here the band metric is six-band STFT-frame EDC MAE:

Output conditionSix-band STFT-EDC error (dB)Broadband EDC error (dB)DRR error (dB)
Original single sample5.3772.6651.821
New broadband head plus projection5.3532.7482.015
Six-band head plus band projection3.0112.9512.258

Octave-band supervision reduces six-band STFT-EDC error by 44.0%, from 5.377 to 3.011 dB, while trading away broadband EDC from 2.665 to 2.951 dB and DRR from 1.821 to 2.258 dB under post-hoc projection with a frozen broadband renderer. Training a multiband-conditioned renderer from scratch is a separate architectural change; we have included this six-band pilot as an ablation in Appendix C.15 (Table 21) to show both the benefit of frequency-dependent planning and the need for joint multiband renderer conditioning.

Reviewer_y25tOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PHILAC, a framework for acoustic field synthesis from unlocalized blind recordings. It combines geometry-aware acoustic decay prediction with discrete diffusion to generate RIRs at unobserved locations without requiring dedicated acoustic measurements.

Critical strengths
  1. Interesting and timely task. Blind acoustic field reconstruction is a novel and practically meaningful problem that removes the need for dedicated RIR measurements and calibrated source positions.

  2. Novel method. The combination of acoustic decay planning and discrete diffusion provides an interesting approach to RIR synthesis.

  3. Promising results. The quantitative improvements are encouraging, and I appreciate the qualitative visualizations of the reconstructed acoustic fields.

Critical weaknesses
  1. Method clarity. Some mathematical formulations, particularly Equation (2), are not sufficiently clear. A more detailed explanation of the variables and their relationships would improve readability.

  2. Related work. The discussion of related work could be more comprehensive. In particular, AVR and AV-Cloud (Both from Neurips 2024) are not adequately discussed, and some comparisons in Table 1 would benefit from further explanation.

  3. Limited perceptual evaluation. The evaluation relies primarily on objective acoustic metrics. A subjective listening study would help demonstrate whether the improvements translate into better perceptual quality and spatial realism.

Questions for authors

I think it would make the paper even stronger that if you can directly record some wet sound in the real environment, and then demonstrate your method effective.

1 comment

We appreciate the encouraging evaluation and constructive suggestions. We are glad that you found the problem timely, the combination of decay planning and discrete diffusion novel, and the quantitative and qualitative results promising. Below we address W1–W3 and Q1 point by point.


W1: Method clarity and variable definitions in Equation (2)

We agree that Equation (2) was overly compact and introduced symbols before defining them explicitly in the surrounding text. In the revised manuscript, we have streamlined Section 3.2 (Eq. 2) and provided a clear step-by-step variable map at first use.


W2: Comprehensive related work (AVR, AV-Cloud) and Table 1 explanations

Thank you for pointing out these two important NeurIPS 2024 works. We have added AVR (Lan et al., NeurIPS 2024) and AV-Cloud (Chen & Shlizerman, NeurIPS 2024) to both Section 2 (Neural acoustic fields) and Table 1, and clarified the taxonomy columns in the Table 1 caption:

  • AVR (Lan et al., NeurIPS 2024): Introduces acoustic volume rendering in the frequency domain with spherical integration to enforce wave-propagation consistency across poses. Like NAF and INRAS, AVR relies on calibrated measured RIRs and per-scene optimization, rather than cross-room zero-shot synthesis from unlocalized blind recordings.
  • AV-Cloud (Chen & Shlizerman, NeurIPS 2024): Learns a sparse calibrated audio-visual anchor cloud from posed observations and uses Audio-Visual Cloud Splatting to render spatial audio at novel listener viewpoints. Unlike Philac, AV-Cloud performs scene-specific novel-view audio rendering from calibrated audio-visual anchors rather than cross-room RIR synthesis from unpositioned ordinary recordings.
  • Clarifying Table 1 columns: In the revised Table 1 caption, we have explicitly defined each comparison axis: Observation (calibrated probe RIR vs. ordinary blind recording), Query (reconstructing the already-observed path vs. synthesizing an unobserved source path), Geometry (conditioning on explicit 3D/depth/visual scene cues), Transfer (cross-room generalization vs. real-room evaluation without per-scene retraining), and Generative (stochastic vs. deterministic mapping).

W3: Subjective perceptual evaluation

We agree that objective acoustic metrics alone do not verify whether the predicted RIRs yield more natural and spatially faithful audio. To evaluate perceptual quality directly, we conducted a blinded MUSHRA listening study (ITU-R BS.1534) with 12 participants across 12 trials spanning Speech, Music, and RIR stimuli (144 evaluations per condition), comparing Philac, FLAC, EigeNet, and the Hidden Reference on a 0–100 perceptual similarity scale. All 12 participants passed the hidden-reference screening check (rating the Hidden Reference ≥90\ge 90 on every trial):

ConditionMean (95% CI)StdRangeQuality Tier
Hidden Reference95.40±0.4095.40 \pm 0.402.4590–100Excellent
Philac (Ours)85.42±0.9385.42 \pm 0.935.6874–96Excellent
FLAC76.40±1.8776.40 \pm 1.8711.4252–91Good
EigeNet56.83±2.6056.83 \pm 2.6015.9425–85Fair

Philac reaches the Excellent perceptual quality tier (85.42±0.9385.42 \pm 0.93) and outperforms the strongest baseline FLAC (76.40±1.8776.40 \pm 1.87) by +9.02 points with substantially lower variance across listeners and stimuli. An anonymous link to the audio examples is included in the supplementary material.


Q1: Direct demonstration on wet sound recorded in a real environment

We already evaluate zero-shot transfer on real measured rooms from the HAA dataset in Tables 4 and 12, reducing query-weighted EDC error from 5.147 dB to 2.797 dB and room-macro EDC error from 5.392 dB to 3.212 dB without room-specific fine-tuning, where blind context inputs are constructed by convolving real measured RIRs with ordinary speech. At the same time, we fully agree that testing directly on in-situ recorded wet audio captured live in a real environment would provide an even stronger end-to-end validation covering real microphone responses, ambient noise, and source directivity. Because collecting a new spatially registered hardware dataset is not feasible within the rebuttal window, we have explicitly highlighted in-situ live wet-recording evaluation as an important future direction in Appendix D of the revised manuscript.

Comments on the paper

1 comment

We thank the AC and reviewers for the constructive feedback. We have substantially strengthened the paper with stronger motivations, new diagnostic studies, spatial and perceptual evaluations, controlled ablations, and clearer exposition. All revisions are highlighted in blue in the revised manuscript PDF.

  • Diagnostic breakdown of the dual-blind relaxation (52QV). Removing clean context RIRs and context poses simultaneously causes a 2.1× to 4.3× EDC error surge across retrained baselines, motivating our pose-free context aggregation, query-relative geometry anchoring, and two-stage decay-guided completion in Section 3.1 and Table 2.
  • Spatial field structure, occlusion contrast, and trajectory consistency (TGJ2, 52QV). Matched-distance LOS to NLOS tests (Table 8) and multi-partition apartment field visualizations (Figure 8) show that PHILAC recovers sharp occlusion transitions rather than distance-only scaling, while outperforming FLAC under heterogeneous context visibility (Table 9) and dense trajectory steps (Table 10).
  • Subjective perceptual study and stochastic decoding analysis (TGJ2, y25t). In a blinded ITU-R BS.1534 MUSHRA test across 12 participants and 12 trials (Table 11), PHILAC reaches the Excellent tier (85.42 ± 0.93) and outperforms FLAC (76.40 ± 1.87) by 9.02 points, with interactive audio examples available in the paper.
  • Absolute peak gain validation and decay-planning pilots (TGJ2, 52QV). Pre-normalization evaluation confirms that our scalar gain token achieves 2.129 dB peak MAE and places 77.4% of predictions within 3 dB of ground truth (Table 19). We also add six-band octave planner and STFT-EDC projection ablations (Tables 21 and 22) alongside baseline adaptation details (Appendix A.6).
  • Clearer formulation, expanded related work, and streamlined appendices (y25t, 52QV). We define all symbols explicitly around Eq. 2, incorporate AVR and AV-Cloud into Section 2 and Table 1, merge overlapping method and implementation details into Appendix A, and condense metric definitions in Appendix B.

Overall, these additions provide stronger empirical and perceptual evidence that unlocalized everyday recordings paired with target geometry support spatially faithful acoustic field synthesis, while clearly delineating current trade-offs and future directions. We look forward to your feedback and remain happy to address any remaining questions during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.