acceptodds
Under review as a conference paper at ICLR 2027

AURA: An Any-format Unified Representation for Spatial Audio

Abstract

Spatial audio is produced and consumed in a zoo of mutually incompatible formats—channel-based layouts such as 5.1, 7.1 and 7.1.4, and scene-based Ambisonics (FOA, HOA). Existing systems either hard-wire a single input/output pair (e.g. stereo-to-7.1.4 upmixing) or route everything through a fixed, order-limited Ambisonics basis whose angular resolution is capped a priori. We introduce AURA (an Any-format Unified Representation for spatial Audio), which maps any spatial-audio format into a single learnable, format-agnostic soundfield latent z_field and decodes it back into any target format. The latent is kept physically grounded by two inductive biases: geometry conditioning, which tags every channel with its physical direction via spherical-harmonic features and treats it as an element of an unordered set, making the encoder permutation-invariant and shared across formats of differing channel counts; and an Ambisonics physics prior, which regularizes z_field toward the deterministic HOA encoding of the input so the latent adapts to content without collapsing into an opaque channel code. Two capabilities follow: any-to-any format conversion, by encoding an input and decoding to the desired layout; and stereo-to-soundfield generation, where a flow-matching model generates z_field from stereo and renders it to arbitrary formats with mutually consistent outputs. On real corpora (STARSS22/23, Freesound 5.1, 3D-MARCo), the frozen front-end preserves spatial structure almost losslessly (sub-degree to 1.5° round-trip DOA error), and on the real 5.1 test set AURA reproduces the reference's inter-channel energy structure far more faithfully than a deterministic upmix (44.3 dB dynamic range vs. 91.0 dB, reference 57.9 dB).

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Spatial audio, Soundfield representation, Ambisonics, Flow Matching

Reviews

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

This paper proposes AURA, a unified representation for spatial audio that aims to support arbitrary input and output spatial-audio formats. The key idea is to encode multi-channel audio as an unordered set of channel-wise latent representations conditioned on the physical direction of each channel, producing a format-agnostic latent representation zfieldz_{\mathrm{field}}. A geometry-conditioned decoder can then reconstruct the representation in an arbitrary target format. The model additionally incorporates a physics-inspired regularization toward a deterministic HOA representation and a flow-matching diffusion transformer for stereo-to-soundfield generation. The authors evaluate AURA on STARSS22/23, Freesound 5.1, and 3D-MARCo. The experiments include self-reconstruction, cross-format conversion, cross-array conversion, and stereo-to-spatial generation.

Critical strengths
  1. Clean unified representation formulation
    AURA provides a compelling formulation of spatial audio as a geometry-conditioned set representation. Treating channels as an unordered set while explicitly conditioning on their physical directions is a natural and potentially powerful way to decouple the representation from a particular channel ordering or format. This is a meaningful system-level perspective for spatial-audio representation learning.

  2. Practical any-to-any conversion framework
    The same latent representation can be decoded into multiple target spatial formats, providing a unified interface for cross-format conversion. The experiments on 3D-MARCo demonstrate that the approach can handle several nontrivial conversion scenarios, including cross-array conversion. This makes the work practically relevant beyond a single spatial-audio task.

Critical weaknesses

The central “any-format” generalization claim is insufficiently validated. Although the authors claim that AURA can generalize to arbitrary spatial configurations, the current experimental setup does not convincingly demonstrate this capability. In particular, the evaluation does not clearly establish whether the proposed geometry-conditioned representation can generalize to spatial configurations that are unseen during training. Several concerns regarding this claim are raised in the following section.

Questions for authors

Q1. Unseen geometries and datasets.
In Table 1, the authors list the datasets and spatial formats used for training and evaluation. However, the current experiments do not explicitly evaluate generalization to unseen datasets or spatial geometries. Could the authors provide experiments in which certain spatial configurations or geometries are held out from training and used only for evaluation? For example, the authors could remove one or more geometries from the training set and evaluate AURA on these unseen geometries. Such experiments would help distinguish the architectural capability to support arbitrary geometries from empirically demonstrated generalization to unseen spatial configurations.

Q2. Role of geometry conditioning.
In Table 5, replacing geometry conditioning with index embedding results in comparable performance, with some metrics even slightly improving. This raises a question about whether the geometry conditioning is actually necessary for the claimed generalization to arbitrary spatial configurations. Could the authors provide further analysis of the role of geometry conditioning, particularly by evaluating both variants on unseen geometries? Such an analysis would help clarify whether geometry conditioning provides a meaningful advantage for generalization beyond the spatial configurations observed during training.

Q3. Effect of the spherical-harmonic order NshN_{\text{sh}}.
In Table 5, changing NshN_{\text{sh}} from 3 to 2 results in very similar performance, with small enhancements in Reconstruction Spec. SNR and Cross Latent-SNR. Could the authors provide further analysis of the role of NshN_{\text{sh}} in the proposed representation? In particular, it would be helpful to evaluate additional values, such as Nsh=1N_{\text{sh}}=1 and Nsh=4N_{\text{sh}}=4, to better understand how the spherical-harmonic order affects the model's reconstruction and cross-format generalization performance.

3 comments

To Reviewer XXZm

Q1: held-out geometries. Done; the answer is mostly negative, and the paper now says so. We retrained with every Hamasaki 12-channel example removed and decoded to Hamasaki at test time.

9-ch→Hamasakilat-SNR (all 13 / held-out 2)|ΔE| (all / held-out)
AURA, Hamasaki seen1.66 / 1.731.60 / 1.71
AURA, Hamasaki unseen−0.15 / 0.012.80 / 2.84
HOA mode-matching N=3−1.713.06
HOA max-r_E N=4 (best energy)−1.650.45

With the layout unseen, AURA degrades gracefully:

  • It stays above every fixed-HOA variant on content.
  • On energy it is on par with mode-matching but far behind max-r_E.

With six training layouts it does not extrapolate. We agree with your distinction between architectural capability and demonstrated generalization. The paper now claims only the former, and §3.3, §4.5 and the conclusion are revised accordingly.

Q2: SH geometry conditioning vs. index embedding on an unseen geometry. We ran exactly this comparison:

Hamasaki unseenRecon SNRCross lat-SNRCross |ΔE|
SH geometry cond.9.0−0.152.80
Index embedding10.1−0.284.06

The index embedding reconstructs the seen layouts slightly better; it has fewer constraints while memorizing six layouts. However, it transfers worse to the unseen one: |ΔE| is 1.3 dB worse, which is well beyond the ±0.18 dB seed spread. Geometry conditioning therefore helps where it should, on an unseen layout, but modestly. Its main benefit remains structural: permutation invariance, one encoder across channel counts, and a continuous query space.

Q3: SH order N_sh (including 1 and 4). Added (13 scenes, 40k steps):

N_shRecon SNRCross lat-SNRCross |ΔE|
19.61.651.49
211.51.681.80
3 (default, 3 seeds)10.8 ± 0.91.65 ± 0.011.60 ± 0.18
49.81.661.68

All values lie within seed noise. The representation is insensitive to N_sh on seen layouts, so even N_sh = 1 (4 features per channel instead of 16) suffices there. We keep N_sh = 3 as the default because it costs almost nothing at these channel counts. We now report this insensitivity rather than claim an optimum.

Thank you for the detailed rebuttal and the additional experiments. I appreciate the authors' efforts to address my concerns and the clarifications regarding generalization to unseen geometries and the role of geometry conditioning. Nevertheless, my overall assessment remains unchanged, and I will maintain my original score.

Thank you very much for your reply. We will further optimize the algorithm in the future, aiming to enhance its generalization performance in layouts unseen before.

Reviewer_nf4nOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes AURA, a unified learnable representation for spatial audio that supports both any-to-any spatial audio format conversion and stereo-to-spatial audio generation. The main motivation is that existing spatial audio methods either rely on fixed-order Ambisonics (HOA), whose spatial resolution is limited by the chosen order, or employ neural generative models designed for a fixed output channel layout. To address these limitations, AURA introduces a format-agnostic sound field latent representation, learned through a geometry-conditioned Set Transformer encoder and a layout-conditioned decoder. It incorporates an auxiliary HOA prediction loss to encourage physically meaningful spatial representations, together with reconstruction, cross-format, cycle-consistency, and energy-alignment objectives.

For spatial audio generation, the authors employ a Flow Matching model to generate the unified sound field latent from stereo audio, enabling the same generated representation to be rendered into different output formats. Experiments demonstrate promising energy preservation in several cross-format conversion tasks and improved inter-channel energy distribution compared with deterministic stereo upmixing. However, the evidence supporting superior spatial fidelity and generation quality remains limited.

Critical strengths
  1. Well-motivated and meaningful research problem: The paper addresses an important problem in spatial audio representation and generation. The limitations of fixed-format generation and the desire for a unified representation are natural and meaningful motivations.

  2. Reasonable and coherent method design: The proposed geometry-conditioned encoder and decoder provide a natural architecture for handling different channel counts and spatial layouts. The use of an auxiliary HOA physics prior is also reasonable.

  3. Well-structured training objectives and useful ablation studies: The reconstruction, cross-format, latent-consistency, HOA-prior, and energy-alignment losses are reasonably designed for their respective purposes. The authors also conduct ablation studies investigating the contributions of different components and the latent representation capacity.

Critical weaknesses
  1. Missing Important Generative Baseline: The paper does not compare against ImmersiveFlow, its closest related work.

  2. Missing HOA Order Scaling Experiments: The paper argues that fixed-order HOA limits spatial representation capacity, but does not investigate how increasing the HOA order affects conversion quality, spatial fidelity, and computational cost. Moreover, the HOA order used in the main comparison table is not explicitly specified.

  3. Limited Improvements over HOA: AURA does not consistently outperform fixed HOA in format conversion. Furthermore, improvements in energy preservation alone are insufficient to demonstrate superior spatial representation or localization accuracy.

  4. Insufficient Ablation Evidence: Although the paper includes ablation studies, removing geometry conditioning or the HOA prior does not consistently degrade cross-format performance. Their contributions to spatial generalization remain insufficiently validated.

Questions for authors

All questions are addressed in the Weaknesses section.

3 comments

To Reviewer nf4n

W1: missing ImmersiveFlow baseline. Added, under identical data, VAE, backbone, objective, sampler and budget (Table 3):

AURAImmersiveFlow
FAD per-channel / downmix0.311 / 0.3000.492 / 0.350
Per-channel dynamic range (ref 57.9 dB)44.332.2
Mean loudness error17.4 dB23.1 dB
LFE error+16.8 dB+26.0 dB

Because only the generation target differs, this isolates the value of the learned z_field over a fixed-layout target.

W2: HOA order scaling; unstated order. Both are fixed. The main table uses N = 3 mode-matching, now stated in the caption and text. Table 7 sweeps N ∈ {1,2,3,4} across four decoders (mode-matching, SVD, max-r_E, EPAD).

  • Latent-SNR is flat in N for the Hamasaki target: −1.55 to −1.76 dB for every decoder. Raising the order does not help match a channel-based recording. The bottleneck is not angular resolution but that a fixed linear map cannot reproduce how a different array picks up the scene.
  • Energy depends on the decoder, not on order:
    • Mode-matching goes from 1.29 dB (N=1) to 3.06 dB (N=3).
    • max-r_E at N=4 reaches 0.45 dB, better than AURA.
    • The earlier "HOA is much worse on energy" impression was partly an artefact of a single decoder.
  • 9-ch→HOA* row: the target is an order-3 encoding, so N≥3 matches it trivially (+3.25 dB). This row is a sanity check, not a win for either method.
  • Cost: HOA grows as (N+1)² channels. AURA's latent is fixed (12 × 64) regardless of source and target layouts.

W3: AURA does not consistently beat HOA; energy alone is insufficient. We agree, and the claims have been rewritten. What survives:

  1. On content (latent-SNR against the real target), AURA beats fixed HOA at every order and decoder tested, on all and on held-out scenes.
  2. On energy, a tuned max-r_E decoder at N=4 is better than AURA, and the paper says so.
  3. On an unseen geometry, AURA's advantage largely disappears. Its content score is still above HOA, but its energy is only on par with mode-matching.

The paper no longer argues "AURA is more accurate than HOA" as a headline. It argues for one shared, fixed-size representation that covers conversion, cross-array transfer and generation, and that beats a fixed-target generator under identical conditions.

W4: ablations do not support geometry conditioning / the HOA prior. For seen layouts this is correct. Neither effect is distinguishable from seed noise, and we removed that claim. On the unseen layout, which is where geometry conditioning should matter, SH features beat an index embedding on both cross metrics:

| Hamasaki unseen | lat-SNR | |ΔE| | |---|---|---| | SH (ours) | −0.15 | 2.80 dB | | Index | −0.28 | 4.06 dB |

The gap on |ΔE| (1.3 dB) is well beyond the seed spread (±0.18). It is a modest benefit, and we present it as such, not as generalization.

The new results address some of my concerns, particularly regarding the ImmersiveFlow baseline and HOA order scaling. I will raise my score.

Thank you very much for your reply and for your recognition of our work. We will revise the paper based on the key comments from all the reviewers.

Reviewer_ffaUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a unified representation shared across spatial audio formats, including 5.1, 7.1.4, and Ambisonics, so that any format can be converted to any other. Each channel is encoded by a frozen VAE and tagged with its direction, and the resulting channel embeddings are then pooled. The decoder is conditioned on the target loudspeaker directions, so a single model can render to any layout. An auxiliary Ambisonics head serves as a regularizer and also produces the Ambisonics output. In addition, a stereo-to-representation generator enables upmixing. The method is evaluated on format conversion using 3D-MARCo and on stereo-to-5.1 upmixing using 48 clips. The paper includes no comparison with ImmersiveFlow and no listening test.

Critical strengths
  • Layout-agnostic decoding. Because the decoder is conditioned on loudspeaker directions, a single model covers arbitrary layouts and produces mutually consistent outputs. This addresses the fixed-layout limitation of prior upmixers and is of clear practical value.
  • Careful treatment of the fundamentals. The VAE reconstruction ceiling is measured separately, and Ambisonics conventions are applied correctly.
Critical weaknesses
  1. Data split and baseline issues.
  • The STARSS23 development set already contains the entire STARSS22 development set (121 + 47 = 168, per the official documentation). The paper does not use the official splits (173 = 121 + 52 vs. 247 = 168 + 79), so train/test overlap is likely.
  • 3D-MARCo is split 11/2, yet format conversion is evaluated on 13 scenes, which appears to include training scenes.
  • The Ambisonics baseline uses mode-matching (pseudo-inverse) decoding, which is known to produce uneven energy distribution on irregular layouts (Zotter et al., 2012). With the paper’s own nominal angles, the Hamasaki 12-channel layout at 3rd order has rank 11, and the decoder remains ill-conditioned even under small angular errors.
  • The proposed method outperforms the baseline in energy error only in the three cross-layout rows. Energy error is also the quantity targeted by the loudness loss, which is described only in the appendix and is not ablated.
  1. Main claims are not supported by clear evidence.
  • The per-channel VAE does not preserve phase, whereas FOA directional encoding depends on inter-channel phase.
  • The abstract lists STARSS, but Table 2 contains no FOA results.
  • In the ablation study, removing direction tagging or the Ambisonics head yields equal or better cross-format results. The number of seeds and the variance are not reported.
  • There is no comparison with ImmersiveFlow, the most closely related work, on the same data, even though its training code is publicly available. FAD is roughly 3× worse per channel and 1.6× worse on the downmix, and it is computed from only about 8 minutes of audio.
  • Although Sections 3.3 and 3.5 claim generalization to unseen layouts, no such test is provided. Parametric methods such as HO-DirAC and COMPASS are also missing from the comparison.
Questions for authors
  1. Clarifications
  • Several details appear inconsistent: The number of training steps is reported as both 4e5 and 2e5. Which is correct? The upmixing baseline is not described in Section 3.9. Why is it omitted? Section 4.2 uses 48 clips from 3D-MARCo, whereas Table 1 reports 13. What accounts for this difference? The “spec SNR” of 11.3 in Table 5 equals the latent SNR in Table 3. Is the label correct?
  • Do the two datasets (STARSS22 and STARSS23) overlap in your setup? If the overlap was removed, how was this done? Why were the official splits not used?
  • Do the 13 3D-MARCo scenes include training scenes? Can results on held-out scenes be provided? What are the results without the loudness loss, and against an energy-preserving decoder such as AllRAD or EPAD?
  • Can the authors provide a comparison with ImmersiveFlow trained on the same Freesound 5.1 data?
  • The appendix states that the VAE operates at 44.1 kHz / 21.5 fps, whereas the LeVo VAE operates at 48 kHz / 25 Hz. Which configuration was used?
  1. Suggestions
  • Evaluation on unseen layouts, a listening test, and experiments on real stereo masters would substantially strengthen the paper.
2 comments

To Reviewer ffaU (1/2)

W1: STARSS22 ⊂ STARSS23. You are right. Our splits were drawn independently, so they overlap: 13 of our STARSS23 test clips have byte-identical copies in STARSS22 training. Fortunately this does not affect any reported number:

  • STARSS is used only to train the autoencoder.
  • No table reports a STARSS test number.
  • Every conversion and generation result comes from 3D-MARCo or Freesound.

The dataset table now states the overlap. We will move to the official STARSS23 split with deduplication.

The same point answers "the abstract lists STARSS but Table 2 has no FOA result". The revised abstract lists only the corpora that are evaluated. FOA appears only in the front-end analysis (see "FOA phase" below).

W2: 3D-MARCo 11/2 split, but conversion evaluated on 13 scenes. Correct; the 13-scene numbers include training scenes. Table 2 now reports both all-13 and the 2 held-out scenes. On held-out scenes the energy advantage over mode-matching HOA shrinks but survives:

| Pair | AURA |ΔE| | HOA (N=3) |ΔE| | |---|---|---| | 9-ch→Hamasaki | 1.25 | 2.80 | | Cross-array | 2.41 | 3.11 | | Hamasaki→9-ch | 2.80 | 2.78 (tie) |

The content advantage also holds on held-out scenes. AURA reaches 1.4–2.5 dB latent-SNR, while every HOA variant on 9-ch→Hamasaki stays at ≤ −1.4 dB.

W3: mode-matching is ill-conditioned; use AllRAD/EPAD. Agreed. This was the most useful criticism. Table 7 now covers N ∈ {1,2,3,4} × {mode-matching, truncated SVD, max-r_E, EPAD} (13 scenes, 14 s excerpts).

Findings:

  • max-r_E at N=4 is the best decoder on energy. It reaches |ΔE| = 0.45 dB, better than AURA's 1.70 dB on the same excerpts.
  • EPAD behaves as intended. It is direction-uniform by construction: the plane-wave energy spread is ≤ 1.2 dB at N≤2, versus 19 dB for mode-matching.
  • EPAD's |ΔE| against the real Hamasaki recording is large (5.5–6.1 dB). This is a calibration effect, not a defect. On the same scenes the 9-ch recordings are 4.5 ± 0.4 dB louder than the Hamasaki recordings. A decoder that preserves source energy exactly therefore lands about 4.5 dB from the target, whereas AURA learns each array's level from paired data. We now say explicitly that |ΔE| against a real target mixes energy preservation with inter-array gain.
  • On content, no decoder at any order closes the gap. Every variant stays at or below −1.5 dB latent-SNR on 9-ch→Hamasaki (best: −1.55 dB), versus +2.31 dB for AURA. Energy-preserving decoding fixes the level, not the content of each channel.

The paper no longer claims that AURA is uniformly better on energy.

W4: the energy result is what the loudness loss targets, and it was not ablated. Fixed, at both levels.

  • Autoencoder:
    • Setting λ_e = 0 lowers reconstruction from 10.8 to 7.5 dB.
    • It raises cross-format |ΔE| from 1.60 to 2.56 dB (held-out: 1.71 → 2.28).
  • Generation (new column in Table 3): without L_energy, the generated 5.1 has
    • a per-channel dynamic range of 6.8 dB (AURA 44.3; reference 57.9),
    • an LFE +35.0 dB too loud,
    • FAD rising from 0.311/0.300 to 0.579/0.584.

The energy advantage is therefore substantially attributable to L_energy, and §4.5 now says so.

W5: seeds/variance; the ablations do not support the claims.

  • Three seeds of the full model: 10.8 ± 0.9 dB reconstruction, 1.65 ± 0.01 dB cross latent-SNR, 1.60 ± 0.18 dB |ΔE|.
  • Robust effects against this yardstick: only S = 8 slots (6.7 dB) and the loudness term.
  • Within noise on seen layouts: the HOA prior is borderline, and cycle consistency, the index embedding and N_sh are within noise.

We therefore removed the claim that geometry conditioning or the HOA prior improves accuracy on seen layouts. The test where geometry conditioning should matter is an unseen layout. That test is now in the paper (see W7).

To Reviewer ffaU (2/2)

W6: no ImmersiveFlow comparison on the same data; FAD duration. We added it. ImmersiveFlow is retrained on our Freesound 5.1 training set, with the same frozen VAE, DiT (1024-d, 12 layers, 16 heads), OT-CFM objective, 50-step Euler sampler, 2×10⁵ steps and seed as AURA. The only difference is the target: a learned z_field rendered to 5.1, versus a fixed 6×64 latent tensor.

(48 test clips)AURAImmersiveFlowPADRef.
FAD per-ch / downmix ↓0.311 / 0.3000.492 / 0.3500.385 / 0.214—
Dynamic range (dB, →ref)44.332.228.357.9
Mean |ΔdB| vs. ref ↓17.423.130.6—
LFE vs. ref (dB, →0)+16.8+26.0+44.9—

Under identical conditions, AURA is better than ImmersiveFlow on both FAD aggregations and on every loudness statistic.

The "3× / 1.6× worse FAD" you note is relative to the passive upmix (0.103 / 0.180). That upmix copies the input stereo almost verbatim into L/R and leaves the other channels near-silent, so its spectrum trivially matches the reference distribution. Its level structure is badly wrong, however: dynamic range is 91.0 dB against a reference of 57.9, and the LFE is −34.4 dB. We therefore report FAD together with loudness and do not claim FAD superiority over training-free upmixes.

On duration, our original text was misleading. The "≈10 s" described training crops, not test clips. The test set is 48 full-length clips (≈56 min). The DiT's positional table caps a clip at 185 s, which truncates 4 of the 48 clips and leaves ≈39 min scored. This is now stated in §4.4 and the table caption. FAD also varies by ≈0.02 between runs on identical audio; we state this and do not read smaller gaps as meaningful.

W7: no unseen-layout test; HO-DirAC/COMPASS.

  • Unseen layout: see the reply to XXZm Q1/Q2 for full numbers. In short, with Hamasaki held out of training:

    • latent-SNR falls to −0.15 dB and |ΔE| rises to 2.80 dB;
    • this remains above every fixed-HOA variant on content;
    • on energy it is only on par with mode-matching and well behind max-r_E.

    With six training layouts, AURA does not extrapolate. §3.3 has been rewritten accordingly, and the conclusion lists this as a limitation.

  • Parametric methods: HO-DirAC and COMPASS are now discussed in related work; a faithful reimplementation was not feasible in the rebuttal window, so we list it as a needed baseline rather than claim a comparison.

Clarifications (Q1).

  • Steps: both the autoencoder and the generator train for 2×10⁵ steps; "4×10⁵" was an error. Ablations use a stated 4×10⁴-step budget.

  • Upmix baseline: this was under-described. §3.9 now specifies both upmixes:

    • The passive upmix encodes stereo to HOA and decodes to 5.1. L/R stay nearly unchanged; C, LFE and the surrounds are near-silent.
    • The PAD upmix works per STFT bin. It uses a coherence ambience mask sent to the surrounds with a 10 ms delay, a similarity-gated centre, and a 120 Hz LFE.

    Against PAD, AURA wins on per-channel FAD (0.311 vs 0.385) and on loudness (17.4 vs 30.6 dB error). PAD wins on downmix FAD (0.214 vs 0.300), because its downmix is essentially the input.

  • 48 clips vs. 13 scenes: both counts are correct but refer to different units. The 13 are 3D-MARCo scenes. The 48 in the front-end study are 9-channel array recordings: 13 scenes × 4 arrays, minus missing takes. The text now says "array recordings".

  • "Spec SNR 11.3 = latent SNR 11.3": the labels were correct, and the match is a coincidence of rounding between two different measurements:

    • Ablation table: STFT-magnitude SNR after VAE decoding, 11.32 dB.
    • Conversion table: latent-domain SNR, 11.26 dB.

    The two even come from different evaluation windows. With three seeds the ablation value is now 10.8 ± 0.9 dB, and the ablation caption explicitly distinguishes the two metrics.

  • VAE: the VAE is Stable Audio Open's (44.1 kHz, ×2048 downsampling, 21.53 fps). "LevoVAE" was a stale code label, now removed from the paper. LeVo's weights were not available to us, which is also why ImmersiveFlow had to be retrained.

FOA phase. Correct, and this is a real limitation. Per-channel encoding discards the inter-channel phase on which first-order directional encoding depends. After the front-end round-trip, FOA DOA error is 16.0° on average (median 8.7°). Loudspeaker formats, by contrast, stay at 0.3–1.5°. We make no FOA directional claim. The abstract and conclusion state the limitation, and a phase-aware multichannel front-end is the first item of future work.

Suggestions. Unseen layouts: done. Real stereo masters (our condition is a downmix of the 5.1 master): future work. Listening test: not run; protocol in App. A.8.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.