REVIEW 3 major objections 5 minor 33 references
Room Impulse Response Generation Conditioned on Acoustic Parameters
T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read A model conditioned only on acoustic parameters can generate room impulse responses that match or beat geometry-conditioned baselines.
desk verdict Solid empirical comparison of DAC-based generative models for acoustic-parameter-conditioned RIR generation; MaskGIT wins, but a Motus train/test overlap ambiguity and reference-spectrum postprocessing need fixing before the headline claim is credible. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the Descript Audio Codec (DAC) codegram, a $9 \times 172$ grid of discrete tokens per two-second RIR, produced by nine residual vector-quantization codebooks at a frame rate of 86 Hz; the same representation can also be read as continuous latent vectors. The MaskGIT-style variant (MGALN) carries the argument: it sums the projected embeddings of the codebooks and injects the acoustic-parameter conditioning into each transformer layer through adaptive layer normalization (adaLN), then decodes tokens with masked parallel sampling and classifier-free guidance. Cross-attention and classifier-guidance variants of the autoregressive transformer, and the flow-matching model on continuous latents, serve as comparators showing which parts of the conditioning and decoding machinery matter.
What would settle it
Retrain MGALN with those 50 test responses held out and re-run the parameter-error and MUSHRA comparisons; if the errors on $T_{30}$, $C_{80}$, and the listener ratings do not remain smaller than StoRIR and FastRIR, the claimed advantage is due to overlap. Concretely, the test responses have known parameter values, so the objective metrics alone can be recomputed on a model trained on the remaining 78,127 samples.
Extended reading notes
Core claim
The paper's central claim is that conditioning on broadband and bandwise acoustic parameters—rather than on room geometry—is enough for a generative model to reproduce a room's response. The evidence is a controlled comparison in which all four proposed models use the Descript Audio Codec as their token or latent space and the same equalization postprocessing, while each baseline is evaluated with its own required conditioning inputs. MGALN, the MaskGIT-based non-autoregressive model with adaptive layer normalization, is reported as the best overall: it has the smallest A-weighted relative errors for $T_{30}$, $T_{15}$, EDT, $D_{50}$, and $C_{80}$, the smallest SRMR deviation, and an average MUSHRA score of 70, corresponding to "good" on the MUSHRA scale. The conclusion the paper draws is that acoustic-parameter conditioning is a viable and perceptually relevant alternative to geometric conditioning, and that the masked-token non-autoregressive architecture is the strongest of the four tested.
Load-bearing premise
The load-bearing premise is that the 50 real room impulse responses used for testing were excluded from the 78,177 training samples; the paper lists the same dataset among training sources yet never states that the test subset was held out.
Editorial extensions
If this is right
- A user can specify reverb time, clarity, definition, and source-receiver distance and obtain a full impulse response without any room geometry, enabling reverb design for imagined or inaccessible spaces.
- Because the conditioning is acoustic rather than geometric, generated responses can be perceptually realistic without corresponding to any real room.
- Non-autoregressive masked generation at the DAC frame rate gives fast inference, suitable for interactive augmented reality, virtual reality, and postproduction workflows.
- The small measured degradation of the DAC encode/decode loop relative to reference RIRs indicates that token-domain generation does not itself limit fidelity.
- If MGALN's results hold, acoustic-parameter conditioning provides a simpler interface than geometry-conditioned pipelines and a stronger baseline than statistical reverb synthesis.
Reading between the lines
- The paper does not state whether the 50 real test RIRs were excluded from the 78,177 training samples even though the same dataset appears in the training list; if they were not held out, the reported gains could reflect memorization rather than generalization.
- The same conditioning vocabulary could support interpolation between measured responses: given two sets of parameters, a model like MGALN could generate a continuum of intermediate reverbs, extending the paper's setting without additional geometry.
- The large source-receiver distance errors that had no audible consequence suggest that distance could be omitted or replaced by a perceptual proxy such as direct-to-reverberant ratio.
- Classifier guidance failed on the autoregressive model; a version of the classifier trained on complete rather than partial RIRs, or applied to the masked model, might rescue conditioning-at-inference-time.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes and evaluates four neural room impulse response (RIR) generators conditioned on acoustic parameters rather than geometry: ARXA (autoregressive transformer with cross-attention), ARCG (autoregressive transformer with classifier guidance), MGALN (MaskGIT with adaLN), and FMIC (flow matching with in-context conditioning). All models operate on Descript Audio Codec (DAC) tokens or latents, and generated RIRs are postprocessed with a reference-derived mel-spectral equalization profile. The models are compared against StoRIR and FastRIR using objective relative errors on the conditioned acoustic parameters plus SRMR, and a MUSHRA subjective test. The paper reports that MGALN achieves the best overall performance, closely followed by FMIC, and that the proposed models outperform both baselines on most metrics.
Significance. The paper's contribution is a systematic empirical comparison of AR versus non-AR generative models and classifier-free versus classifier guidance conditioning for RIR synthesis in the DAC token domain, a task where most prior work conditions on geometry. If the held-out condition for the Motus test set is confirmed and the contribution of the reference-derived mel EQ postprocessing is quantified, the result that a MaskGIT-based model with adaLN can outperform StoRIR and FastRIR would be a useful step toward perceptually driven RIR generation. The paper ships several strengths: a multi-model comparison, bootstrap confidence intervals on all objective metrics, a blind MUSHRA test, and publicly available listening examples. However, the central empirical claim currently rests on two assumptions that are not yet established in the manuscript: strict separation of the Motus test RIRs from training, and the role of the reference-derived mel EQ postprocessing in the reported scores.
major comments (3)
- [Section 4.1 (Dataset)] The test set comprises 50 real RIRs from Motus, and Motus is listed as one of the sources of real training RIRs; the paper does not explicitly state that the 50 test RIRs were excluded from the 78,177 training samples. The sentence 'selecting only one RIR per room or furniture setup to ensure diversity' does not imply a temporal or identity-based split, and the same physical RIR could appear in both the training and test sets. If the test RIRs are included in training, every objective and subjective result in Table 1 and Fig. 2 is inflated by memorization, and the relative ranking of MGALN over the baselines becomes unreliable. Please state whether the test RIRs were held out and describe the split procedure (e.g., file IDs, split by room, split by session), or re-run the evaluation on a disjoint held-out set.
- [Section 2 (mel EQ) and Section 4.5 (Table 1)] The paper states that the system optionally accepts 20 mel-band energy parameters that are used as a postprocessing EQ profile applied to the generated IR, and Section 4.5 states that the same equalization procedure was applied across models. Since the EQ profile is derived from the reference RIR ('ensuring that the timbre of the IR matches the desired spectral characteristics'), every generated RIR receives the target spectral envelope regardless of the generative model's output. This directly affects the bandwise objective metrics in Table 1 (δT30, δT15, δEDT, δD50, δC80) and could also affect MUSHRA judgments of reverberation match. To support the claim that the models themselves outperform the baselines, please report results without the EQ postprocessing or quantify the effect of the EQ (e.g., by applying it to a trivial model or to the Ref-DAC condition). Also clarify whether StoRIR and FastRIR outputs were EQ'd with the same reference-derived profile, and whether the EQ parameters are available to those baselines in practice.
- [Section 4.4 (MUSHRA)] The MUSHRA section lacks key details required by ITU-R BS.1534: the screening procedure and hearing acuity of the 15 participants, the number of trials per condition, the playback system and listening level, whether the hidden reference was correctly identified, and the exact rating scale used (the paper says 'how closely each test sample's reverberation matched that of the reference' rather than the standard MUSHRA quality scale). The paper reports only that 'all pairwise comparisons are statistically significant after correction for multiple comparisons, except for the comparisons between ARXA and FastRIR and between ARXA and FMIC' without giving the test statistic, the multiple-comparison correction, or the number of listeners whose data were kept. Because this is the only independent evidence against the circularity of the objective metrics, the subjective evaluation needs to be described and reported fully.
minor comments (5)
- [Section 1 (Introduction)] The sentence 'most prior methods generating RIRs from acoustic parameters are not based on generative models' is immediately followed by a mention of IR-GAN, which is a GAN and thus a generative model; please rephrase to avoid the contradiction.
- [Table 1 and Section 5 (Conclusion)] The conclusion states that MGALN reached an average MUSHRA score of 70, corresponding to 'good' quality on the MUSHRA scale, but Table 1 reports ΔMUSHRA (a deviation from the reference, with lower values better) as 30 for MGALN. Please define ΔMUSHRA explicitly and reconcile it with the reported score of 70.
- [Section 4.5 (Evaluation results)] The statement that SRD 'does not have any relevance in the resulting output quality' raises the question of why SRD is included as a conditioning parameter and as an evaluation metric; if it is irrelevant to perceptual quality, its role in the conditioning setup should be explained.
- [Section 4.2 (Implementation details)] The hyperparameter notation is inconsistent: for ARCG, 'wk = 1/√3Nb for reverberation times' and 'wk = 1/√2Nb for C80 and D50' are given, but the meaning of Nb (total number of frequency bands including broadband) is not stated until the end of the sentence; consider defining Nb before the formulas.
- [Figure 2 (Boxplot)] The boxplot displays each dot as a listener's average rating; with 15 participants, this may hide per-stimulus variability. Please show per-stimulus or per-listener distributions, or at least state the number of data points behind each box.
Circularity Check
No circularity: objective metrics reconstruct the conditioning labels, but MUSHRA is an independent benchmark and no derivation is claimed.
full rationale
This paper makes no first-principles derivation; it is an empirical comparison of generative models. The conditioning inputs (T30, T15, EDT, C80, D50, SRD) are the same quantities used in the objective metrics (Eq. 3), so high objective scores partly reflect the model's ability to embed its conditioning labels into the generated RIR. The paper itself labels this 'reconstruct acoustic parameters' (Sec. 4.3), not as an independent prediction, and the subjective MUSHRA test with hidden reference and anchor provides an external perceptual benchmark that is not definitionally tied to the conditioning labels. The only citation involving a co-author is [21], used to specify the MGALN sampling recipe; it is not load-bearing for the central comparison. The unaddressed possibility that the 50 Motus test RIRs were not held out from the 78,177 training samples is a data-leakage validity concern, not a circularity of the derivation chain; if present it would inflate results by memorization but would not make the conclusion equivalent to its inputs by construction. Therefore no circular step is identified.
Assumptions & free parameters
free parameters (4)
- Acoustic parameter discretization ranges and class counts =
T30/T15/EDT: 15 classes, 0.1 to 1.5 s; C80: 11 classes, 0 to 20 dB; D50: 13 classes, 40% to 100%; SRD: 10 log classes…
- Classifier guidance weights and lambda =
lambda=1, wk=1/sqrt(3Nb) for reverb times, 1/sqrt(2Nb) for C80/D50, 1 for SRD
- Inference sampling settings =
temperature 0.5 (ARCG), 1.0 (ARXA and MGALN); 20 MGALN steps, 25 FMIC steps
- Mel EQ band count =
20 mel-spaced bands, 20 Hz to 20 kHz
assumptions (4)
- domain assumption The chosen acoustic parameters (T30, T15, EDT, C80, D50, SRD, broadband and bandwise) plus the mel EQ profile are sufficient to capture perceptually relevant reverberation.
- domain assumption The modified Schroeder parameter estimator gives reliable reference values on both real and generated RIRs.
- domain assumption The pretrained Descript Audio Codec represents RIRs without degrading perceptually relevant information.
- domain assumption The 50 Motus test RIRs were not part of the training set.
Cite this review
Pith. "Pith review of Room Impulse Response Generation Conditioned on Acoustic Parameters." pith.science (2026). https://pith.science/paper/5KA4AH42
@misc{pith2026250712136,
author = {Pith},
title = {Pith review of: Room Impulse Response Generation Conditioned on Acoustic Parameters},
year = {2026},
howpublished = {\url{https://pith.science/paper/5KA4AH42}},
note = {Machine review of arXiv:2507.12136}
}
read the original abstract
The generation of room impulse responses (RIRs) using deep neural networks has attracted growing research interest due to its applications in virtual and augmented reality, audio postproduction, and related fields. Most existing approaches condition generative models on physical descriptions of a room, such as its size, shape, and surface materials. However, this reliance on geometric information limits their usability in scenarios where the room layout is unknown or when perceptual realism (how a space sounds to a listener) is more important than strict physical accuracy. In this study, we propose an alternative strategy: conditioning RIR generation directly on a set of RIR acoustic parameters. These parameters include various measures of reverberation time and direct sound to reverberation ratio, both broadband and bandwise. By specifying how the space should sound instead of how it should look, our method enables more flexible and perceptually driven RIR generation. We explore both autoregressive and non-autoregressive generative models operating in the Descript Audio Codec domain, using either discrete token sequences or continuous embeddings. Specifically, we have selected four models to evaluate: an autoregressive transformer, the MaskGIT model, a flow matching model, and a classifier-based approach. Objective and subjective evaluations are performed to compare these methods with state-of-the-art alternatives. Results show that the proposed models match or outperform state-of-the-art alternatives, with the MaskGIT model achieving the best performance.
Reference graph
Works this paper leans on
-
[1]
MESH2IR: Neural acoustic impulse response generator for complex 3d scenes,
A. Ratnarajah, Z. Tang, R. Aralikatti, and D. Manocha, “MESH2IR: Neural acoustic impulse response generator for complex 3d scenes,” in Proceedings of the 30th ACM International Conference on Multimedia , 2022, pp. 924–933
work page 2022
-
[2]
Listen2scene: Interactive material-aware binaural sound propagation for reconstructed 3d scenes,
A. Ratnarajah and D. Manocha, “Listen2scene: Interactive material-aware binaural sound propagation for reconstructed 3d scenes,” in 2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR) . IEEE, 2024, pp. 254–264
work page 2024
-
[3]
Few-shot audio- visual learning of environment acoustics,
S. Majumder, C. Chen, Z. Al-Halah, and K. Grauman, “Few-shot audio- visual learning of environment acoustics,”Advances in Neural Information Processing Systems, vol. 35, pp. 2522–2536, 2022
work page 2022
-
[4]
Novel view acoustic parameter estimation,
R. Falcon-Perez, R. Gao, G. Mueckl, S. V . A. Gari, and I. Ananthabhotla, “Novel view acoustic parameter estimation,” arXiv:2410.23523, 2024
arXiv 2024
-
[5]
FAST-RIR: Fast neural diffuse room impulse response generator,
A. Ratnarajah, S.-X. Zhang, M. Yu, Z. Tang, D. Manocha, and D. Yu, “FAST-RIR: Fast neural diffuse room impulse response generator,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 571–575
work page 2022
-
[6]
RoomFuser: Room Impulse Response Generation using Neural Diffusion Models,
E. Grinstein and Z. Chen, “RoomFuser: Room Impulse Response Generation using Neural Diffusion Models,” Oct. 2023. [Online]. Available: https://github.com/egrinstein/roomfuser
work page 2023
-
[7]
Diffusion- RIR: Room impulse response interpolation using diffusion models,
S. D. Torre, M. Pezzoli, F. Antonacci, and S. Gannot, “Diffusion- RIR: Room impulse response interpolation using diffusion models,” arXiv:2504.20625, 2025
arXiv 2025
-
[8]
StoRIR: Stochastic room impulse response generation for audio data augmentation,
P. Masztalski, M. Matuszewski, K. Piaskowski, and M. Romaniuk, “StoRIR: Stochastic room impulse response generation for audio data augmentation,” arXiv:2008.07231, 2020
arXiv 2008
Show all 33 references
-
[9]
Late reverberation synthesis using filtered velvet noise,
V . V¨alim¨aki, B. Holm-Rasmussen, B. Alary, and H.-M. Lehtonen, “Late reverberation synthesis using filtered velvet noise,” Applied Sciences , vol. 7, no. 5, p. 483, 2017
2017
-
[10]
IR-GAN: Room impulse response generator for far-field speech recognition,
A. Ratnarajah, Z. Tang, and D. Manocha, “IR-GAN: Room impulse response generator for far-field speech recognition,” arXiv:2010.13219, 2020
2010 arXiv
-
[11]
High- fidelity audio compression with improved rvqgan,
R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Kumar, “High- fidelity audio compression with improved rvqgan,” Advances in Neural Information Processing Systems , vol. 36, 2024
2024
-
[12]
Scalable diffusion models with transformers,
W. Peebles and S. Xie, “Scalable diffusion models with transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4195–4205
2023
-
[13]
Diffusion models beat GANs on image synthesis,
P. Dhariwal and A. Nichol, “Diffusion models beat GANs on image synthesis,” arXiv:2105.05233, 2021
2021 arXiv
-
[14]
Kuttruff, Room acoustics
H. Kuttruff, Room acoustics. CRC Press, 2016
2016
-
[15]
Classifier-free diffusion guidance,
J. Ho and T. Salimans, “Classifier-free diffusion guidance,” arXiv:2207.12598, 2022
2022 arXiv
-
[16]
Muse: Text-to- image generation via masked generative transformers,
H. Chang, H. Zhang, J. Barber, A. Maschinot, J. Lezama, L. Jiang, M.-H. Yang, K. Murphy, W. T. Freeman, M. Rubinstein et al., “Muse: Text-to- image generation via masked generative transformers,” arXiv:2301.00704, 2023
2023 arXiv
-
[17]
Dieleman
S. Dieleman. (2022) Guidance: a cheat code for diffusion models. [Online]. Available: https://sander.ai/2022/05/26/guidance.html
2022
-
[18]
Combining multiple probability predictions using a simple logit model,
V . A. Satop¨a¨a, J. Baron, D. P. Foster, B. A. Mellers, P. E. Tetlock, and L. H. Ungar, “Combining multiple probability predictions using a simple logit model,” International Journal of Forecasting , vol. 30, no. 2, pp. 344–356, 2014
2014
-
[19]
Attention is all you need,
A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS’17), 2017
2017
-
[20]
MaskGIT: Masked generative image transformer,
H. Chang, H. Zhang, L. Jiang, C. Liu, and W. T. Freeman, “MaskGIT: Masked generative image transformer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 11 315–11 325
2022
-
[21]
Masked generative video- to-audio transformers with enhanced synchronicity,
S. Pascual, C. Yeh, I. Tsiamas, and J. Serra, “Masked generative video- to-audio transformers with enhanced synchronicity,” arxiv:2407.10387, 2024
2024 arXiv
-
[22]
Flow straight and fast: Learning to generate and transfer data with rectified flow,
Q. L. Xingchao Liu, Chengyue Gong, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” arXiv:2209.03003, 2022
2022 arXiv
-
[23]
LAFMA: A latent flow matching model for text-to-audio generation,
W. Guan, K. Wang, W. Zhou, Y . Wang, F. Deng, H. Wang, L. Li, Q. Hong, and Y . Qin, “LAFMA: A latent flow matching model for text-to-audio generation,” arXiv:2406.08203, 2024
2024 arXiv
-
[24]
Guided flows for generative modeling and decision making,
Q. Zheng, M. Le, N. Shaul, Y . Lipman, A. Grover, and R. T. Q. Chen, “Guided flows for generative modeling and decision making,” arXiv:2311.13443, 2023
2023 arXiv
-
[25]
Dataset of impulse responses from variable acoustics room Arni at Aalto Acoustic Labs ,
P. Karolina, S. J. Schlecht, and V . V ¨alim¨aki, “Dataset of impulse responses from variable acoustics room Arni at Aalto Acoustic Labs ,” Aug. 2022. [Online]. Available: https://doi.org/10.5281/zenodo.6985104
2022 doi
-
[26]
A dataset of higher-order Ambisonic room impulse responses and 3D models measured in a room with varying furniture,
G. G ¨otz, S. J. Schlecht, and V . Pulkki, “A dataset of higher-order Ambisonic room impulse responses and 3D models measured in a room with varying furniture,” in 2021 Immersive and 3D Audio: from Architecture to Automotive (I3DA) , 2021, pp. 1–8
2021
-
[27]
dEchorate: a calibrated room impulse response dataset for echo-aware signal processing,
D. Di Carlo, P. Tandeitnik, C. Foy, N. Bertin, A. Deleforge, and S. Gannot, “dEchorate: a calibrated room impulse response dataset for echo-aware signal processing,” EURASIP Journal on Audio, Speech, and Music Processing, vol. 2021, no. 1, pp. 1–15, 2021
2021
-
[28]
Openair: An interactive auralization web resource and database,
D. T. Murphy and S. Shelley, “Openair: An interactive auralization web resource and database,” in Audio Engineering Society Convention 129 . Audio Engineering Society, 2010
2010
-
[29]
GW A: A large geometric-wave acoustic dataset for audio processing,
Z. Tang, R. Aralikatti, A. Ratnarajah, and D. Manocha, “GW A: A large geometric-wave acoustic dataset for audio processing,” in Special Interest Group on Computer Graphics and Interactive Techniques Conference Proceedings (SIGGRAPH ’22 Conference Proceedings) , 2022. [Online]....
2022
-
[30]
Decoupled weight decay regularization,
I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv:1711.05101, 2017
2017 arXiv
-
[31]
ISO 226:2023 Acoustics — Normal equal-loudness-level contours,
International Organization for Standardization, “ISO 226:2023 Acoustics — Normal equal-loudness-level contours,” https://www.iso.org/standard/ 83117.html, 2023, accessed: 2025-05-07
2023
-
[32]
A non-intrusive quality and intelligibility measure of reverberant and dereverberated speech,
T. H. Falk, C. Zheng, and W.-Y . Chan, “A non-intrusive quality and intelligibility measure of reverberant and dereverberated speech,” IEEE Transactions on Audio, Speech, and Language Processing , vol. 18, no. 7, pp. 1766–1774, 2010
2010
-
[33]
Method for the subjective assessment of intermediate quality level of audio systems,
International Telecommunication Union, “Method for the subjective assessment of intermediate quality level of audio systems,” Recommendation ITU-R BS.1534-1, 2001, also known as MUSHRA (MUltiple Stimuli with Hidden Reference and Anchor). [Online]. Available: https://www.itu.in...
2001
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.