acceptodds
Under review as a conference paper at ICLR 2027

Preference-Optimised Music Mastering under Degradation Chains

Abstract

Music reaches listeners through production, re-recording and distribution stages, and degradations accumulate along the way, several at once and in order. SonicMaster, the prior model for joint mastering and restoration, is trained mostly on single degradations and never on more than three, so it can succeed by correcting each artifact in isolation. We instead train on degradation chains of one to eight degradations (mean 4.8, against 1.6) from 24 types, most of them drawn from six production scenarios, plus pass-through pairs that teach the model when to leave well-mastered music unchanged. On these chains we train Prism, an instruction-conditioned rectified flow from degraded to reference-master latents. Matching a reference alone does not ensure high production quality, so we refine it through a single offline round of preference optimization, ranking the model's own samples by predicted production quality (PQ) and mel-SSIM to the reference. On held-out chains, Prism improves mel-SSIM by versus and better preserves already-mastered inputs (0.98 vs. 0.82). Across 574 real low-production-quality releases it raises predicted PQ from to , improving 79.3% of tracks, against for SonicMaster and for the same architecture trained further on its original corpus. Ablations show that a model trained on the prior corpus barely improves on our chains, and that our advantage holds at every chain depth tested (one to five).

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
music mastering, flow matching, preference optimization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_kGziOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of joint music restoration and mastering under complex, sequential degradation chains. Prior models (such as SonicMaster) were trained primarily on isolated or shallow degradations (at most three independent effects), failing to generalize to compound real-world degradations where multiple flaws co-occur and interact. The authors propose PRISM, an instruction-guided rectified flow framework trained in a frozen autoencoder latent space with FLAN-T5 text conditioning. The contributions include:

  1. A synthesis pipeline combining 24 degradation types (including 8 new realistic mastering and distribution flaws) into ordered chains of 1 to 8 operations across 6 realistic production scenarios, complemented by 15% pass-through pairs to teach non-intervention on clean audio. 2) Mastering Preference Optimisation (MPO), an offline alignment stage ranking perturbed latent rollouts using an equal blend of reference-free Audiobox Production Quality (PQ) and reference mel-SSIM, reinforced with a preferred-sample likelihood anchor. 3) Probing experiments demonstrating that the chain-trained model learns internal representations of acoustic damage that generalize across corpora. Validation spans synthetic chains, a collection of 574 real low-production-quality releases (WILD), and a 10-expert blind listening test. PRISM achieves substantial gains on synthetic compound chains (mel-SSIM +0.112 vs. +0.002 for SonicMaster) and preserves mastered inputs (0.978 vs. 0.821). On WILD, it improves predicted PQ by +0.509, outperforming baselines under loudness-matched conditions, though real-world perceptual superiority remains tied with SonicMaster in a small-scale listening trial.
Critical strengths
  1. Realistic Compound Degradation Modeling with Essential Pass-Through Preservation: By formulating degradation as deep, order-dependent chains representing real-world recording and distribution scenarios, PRISM solves a significant domain-mismatch problem overlooked by prior work. Crucially, incorporating 15% pass-through training allows the model to preserve already-mastered audio (mel-SSIM 0.978 vs 0.821 for SonicMaster), which is vital for any real-world production tool where false-positive alterations ruin professional mixes.
Critical weaknesses
  1. Unvalidated Perceptual Impact of MPO and Weak Instruction Disentanglement: The listening study did not include PRISM w/o MPO, making it impossible to ascertain whether the MPO alignment stage yields human-perceptible audio improvements over the base supervised model. Additionally, instruction sensitivity analysis (Table 3) reveals that natural language prompts only provide coarse on/off intervention gating rather than fine-grained control over which specific co-occurring degradation to repair.
  2. Inconclusive Human Perceptual Evidence on Real-World Releases: While PRISM demonstrates dramatic gains on synthetic chains (85% listener preference), its evaluation on real-world low-quality releases (WILD) in the listening test is limited to only 5 tracks and results in a tie with SonicMaster (42% vs. 42%, 95% CI [16, 68] encompassing chance). Because automated production-quality metrics (such as Audiobox PQ) are known to be vulnerable to reward hacking, the lack of statistically significant human listener preference on real commercial music weakens the claim of real-world mastery transfer.
Questions for authors
  1. In Table 5, on the 5 WILD items, PRISM and SonicMaster tied at 42.0%. Given that SonicMaster boosted loudness by 5.0 LU while PRISM only boosted by 1.0 LU, did listeners in the WILD trials prefer SonicMaster purely due to residual loudness bias, or were there specific artifacts introduced by PRISM? Can the authors conduct a loudness-normalized blind test on an expanded set of WILD tracks (e.g., 20+ tracks)?
  2. In Table 1, MPO improves WILD PQ from 5.807 to 5.873 (+0.066) while slightly degrading pass-through fidelity from 0.984 to 0.978. In informal listening, what qualitative auditory differences distinguish PRISM from PRISM w/o MPO? Did MPO improve high-frequency air, transient punch, or stereo imaging?
  3. Figure 3 shows that linear probing accuracy peaks around block S12-S13. What happens to the audio representations in the final layers (S14-S17) as the flow approaches the reference master? Does the network discard the degradation information to reconstruct the clean target?
3 comments
AuthorsOpenReview ↗

Response to Reviewer kGzi (1/2)

Thank you for recognizing the realistic compound-degradation modeling and the value of pass-through training for preserving mastered audio. Below, we cover the listening evidence, whether MPO's effect is audible, instruction control, and the probing results. Intervals are paired 95% bootstrap intervals unless noted.

W2 and Q1: Listening results on WILD and the role of loudness

Thank you for the suggestion. Here are the per-track votes and loudness for the original five WILD tracks:

ItemAudioSR votesSonicMaster votesPRISM votesSonicMaster minus PRISM loudness
t11145-0.4 LU
t12262+1.7 LU
t13244+6.3 LU
t14262+2.5 LU
t15118+0.8 LU

PRISM was preferred on the two tracks where the outputs were within 1 LU of each other (t11, t15). SonicMaster was preferred only where it was louder (1.7 and 2.5 LU), and even on t13, where it was 6.3 LU louder, PRISM matched it in votes. So loudness likely contributed to the tie, and the votes point to no particular PRISM artifact.

With loudness matched, PRISM keeps a clear lead over SonicMaster on all 574 WILD tracks: +0.345 [0.307, 0.385] PQ. We also ran a new loudness-matched listening test, reported next.

W1 and W2: Does MPO make an audible difference, and does it rely on Audiobox?

We tested this with listeners. In a blind forced-choice test on 15 items (10 from COMPOUND test, 5 from WILD), nine listeners chose between PRISM and PRISM without MPO, with all clips at -23 LUFS. PRISM was chosen in 69 of 135 judgments (51.1%, 95% CI [37.0, 65.2]) and won the majority vote on 8 of 15 items. Listeners chose the two equally often, so MPO's gains in predicted quality and fidelity come at no audible cost.

On the objective side, the WILD improvement holds up once we match each output to its input loudness. PRISM improves PQ over the input by +0.483 [0.432, 0.534], with gains on 80.7% of tracks, and it beats SonicMaster by +0.345 [0.307, 0.385].

We also checked whether MPO's gain depends on Audiobox by retraining MPO with reference mel-SSIM as the only ranking signal, so Audiobox plays no part in training. Using the 2,778 pairs where mel-SSIM picks the same two candidates as the original ranking, we keep 78% of the original MPO gain in loudness-matched WILD PQ: +0.044 [0.040, 0.049], with gains on 88% of tracks. Re-ranking all 10,000 candidate sets by mel-SSIM alone still gives +0.025 [0.021, 0.029], with gains on 71% of tracks.

So most of MPO's PQ gain survives without Audiobox in the ranking, which means the model is not simply reward hacking the metric it is evaluated on.

W1: How much control do instructions give?

Instructions change what PRISM does in two clear ways. First, on already-mastered inputs, a mismatched instruction moves the model further from the input: 1 − mel-SSIM rises from 0.022 to 0.055. The gating effect in Table 3 also grows with MPO, from +0.026 to +0.032, while SonicMaster shows no such pattern (-0.028). Second, on the SonicMaster corpus, swapping the instructions removes 39.0% of PRISM's gain, so the model uses the requested correction rather than ignoring the text.

On compound chains, PRISM repairs the named and unnamed degradations to a similar degree, with median compliance of 0.483 and 0.478 (Appendix K). So when several faults occur together, the model restores the whole chain instead of isolating the one named in the prompt.

We therefore describe the capability as instruction-dependent intervention and sensitivity to the requested correction. Selective control over individual faults within a compound chain is a natural next step, and we leave it to future work.

AuthorsOpenReview ↗

Response to Reviewer kGzi (2/2)

Q2: What changes with MPO?

We compared PRISM with PRISM without MPO on all 574 WILD tracks:

DescriptorPRISM minus PRISM without MPO, with 95% CI
Crest factor+0.122 dB [0.109, 0.136]
Integrated loudness-0.221 LU [-0.235, -0.207]
True peak-0.050 dB [-0.067, -0.033]
Spectral centroid+4 Hz [-2, +11]
Content ceiling-6 Hz [-51, +35]

MPO improves dynamics while keeping the track's tonal character intact. The crest factor increases while loudness and true peak drop slightly, so the output maintains more peak level relative to its average level and is measurably less compressed. Spectral centroid and content ceiling stay the same, which means MPO achieves this without coloring the tonal balance.

In the blind listening test, listeners chose PRISM over PRISM without MPO in 51.1% of 135 judgments (95% CI [37.0, 65.2], chance 50%), so these changes come at no audible cost.

Pass-through fidelity (how unchanged an already-mastered track is left; mel-SSIM between output and input, where 1 means untouched) stays high after MPO: 0.978, compared with 0.984 before, and well above SonicMaster's 0.821.

Q3: What happens to degradation information in the final blocks?

The network does not discard it. On PRISM's own corpus, degradation labels stay readable all the way to the last block, well above chance. What declines after S12–S13 is how well the probes transfer to a different corpus (Appendix P, "Probing analysis: details", Section P.3, Tables 16 and 17):

Probe datasetBest block: mean AP (cAP)S14S15S16S17Chance mean AP
PRISM's own corpusS9: 0.935 (0.90)0.910 (0.80)0.910 (0.78)0.930 (0.90)0.921 (0.89)0.373
SonicMaster test setS12: 0.772 (0.66)0.733 (0.60)0.682 (0.53)0.649 (0.50)0.630 (0.47)0.327

The cAP values for S14–S17 are computed from the per-group AP reported in Tables 16 and 17.

We interpret the intermediate blocks as encoding corpus-invariant degradation features, while later blocks specialize these into the output-specific correction, i.e., the predicted velocity. This is consistent with the training objective v⋆ = z⋆ − z_deg.

Depth in the network and progress along the flow are two different axes, so we also probed along the flow path, using inputs z_t = (1 − t) z_deg + t z⋆ at the best block from Table 18 (Appendix P, "Probing analysis: details", Section P.3):

Flow time t00.250.50.751 (reference)
PRISM mean AP0.9350.9290.9060.7710.380

Chance mean AP is 0.373. Degradation labels remain readable through the intermediate flow states and fall to chance only at the clean reference, as expected, since no degradation is left in the input at that point. In short, the model tracks the damage for as long as there is damage to repair.

Dear Reviewer,

Thank you for your thoughtful feedback. As the discussion period draws to a close, we would appreciate hearing whether our additional experiments and responses address your concerns about MPO’s perceptual benefit, instruction control, and real-world performance, or if further clarification would be helpful.

Best regards,
Authors

Reviewer_YmPaOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PRISM, an instruction-conditioned rectified-flow model for joint music restoration and mastering. It constructs a synthetic corpus with 24 degradation types arranged into production-motivated chains of up to eight operations, together with pass-through examples for preserving already-mastered audio. Starting from SonicMaster, the model is trained on the combined old and new corpora and refined through one offline round of mastering preference optimization (MPO), where candidates are ranked jointly by Audiobox production quality and reference mel-SSIM. PRISM substantially improves fidelity and preservation on held-out synthetic chains, and a small expert listening study strongly favors it on synthetic examples. Results on 574 real releases show higher predicted production quality, but the five-item real-audio listening subset does not distinguish PRISM from SonicMaster.

Critical strengths

1.Meaningful compound-degradation benchmark. The broader degradation taxonomy, production-motivated ordering, deeper chains, and pass-through cases address important limitations of existing music-restoration data. Same-architecture comparisons show that continued training on the old corpus alone does not explain the gains.

2.Strong controlled results. PRISM consistently improves paired and distributional metrics on synthetic chains, preserves mastered inputs, and is strongly preferred by experienced listeners on the synthetic subset. The paper also reports relevant ablations and uncertainty estimates.

Critical weaknesses

1.Real-world and MPO evidence is not independent. Audiobox is involved in reference selection, MPO ranking, WILD selection, and WILD evaluation. The real-audio listening study has only five items, excludes the unprocessed input and PRISM without MPO, and finds no advantage over SonicMaster. The paper therefore does not establish a perceptual benefit on real releases or an audible contribution from MPO.

2.The source of the corpus gain is under-isolated. The new corpus simultaneously changes degradation types, parameter ranges, ordering, chain depth, instructions, and pass-through data. The paper does not separately identify the effects of ordering, new degradation types, and pass-through pairs; the nearly constant advantage across depths also suggests that deeper chains alone are not the main cause.

Questions for authors

1.Do the authors have a larger, loudness-matched WILD listening study that includes the unprocessed input and PRISM without MPO, with labels randomized per listener?

2.When degradation types, severity, and depth are fixed, what is the isolated contribution of production ordering? What are the separate contributions of the new degradation types and pass-through pairs?

5 comments
AuthorsOpenReview ↗

Response to Reviewer YmPa (1/2)

We thank the reviewer for the careful reading and for the constructive suggestions on the real-audio evaluation and on separating the sources of the corpus gains. Unless stated otherwise, all intervals below are paired 95% bootstrap confidence intervals.

W1 and Q1: New listening test with PRISM without MPO

Thank you for the suggestion. We ran a new blind forced-choice listening test that isolates MPO. It covers the same 15 items from the initial submission: 10 from the COMPOUND test set and 5 real releases from WILD. In each trial, listeners chose the better of two versions: PRISM or PRISM without MPO. All clips are 30 s long and normalized to -23 LUFS (BS.1770), so loudness cannot sway the choice. Trial order and the on-screen order of the two options are shuffled separately for each listener.

Nine listeners completed the test, giving 135 judgments. PRISM was chosen in 69 of them (51.1%, 95% CI [37.0, 65.2]). Listeners chose the two equally often, so MPO's +0.056 PQ gain on WILD comes at no audible cost, on top of the +0.289 PQ lead over SonicMaster that PRISM has before MPO. The test does not include SonicMaster; on WILD, PRISM leads it by +0.345 [0.307, 0.385] PQ at matched loudness (table below).

W1: Dependence on Audiobox

Audiobox is never used as a differentiable training loss; its influence on training comes through the preference labels in the original MPO procedure. To measure how much MPO depends on it, we ran two additional experiments that remove Audiobox from candidate ranking entirely. Both start from PRISM without MPO and keep the paper's MPO configuration:

  • E3 ("mel-SSIM agrees" in the revised paper): uses the 2,778 original candidate sets where ranking by reference mel-SSIM alone picks the same preferred and rejected candidates as the original ranking.
  • E3b ("mel-SSIM only"): uses all 10,000 original candidate sets, with the preferred and rejected candidates chosen by reference mel-SSIM alone.

All changes below are relative to PRISM without MPO, and WILD outputs are loudness-matched.

MeasureE3: 2,778 matching pairsE3b: 10,000 re-ranked pairsOriginal MPO: 10,000 pairs
WILD PQ change+0.044 [0.040, 0.049]+0.025 [0.021, 0.029]+0.056 [0.051, 0.062]
WILD tracks with increased PQ88%71%87%
WILD CE change+0.036+0.016+0.052
WILD CU change+0.055+0.034+0.073
Test mel-SSIM change ↑+0.0010 [0.0006, 0.0015]+0.0015 [0.0011, 0.0020]+0.0014 [0.0008, 0.0019]
Test log-spectral distance change, dB ↓-0.071 [-0.100, -0.042]-0.193 [-0.227, -0.159]-0.177 [-0.211, -0.144]
Test PANNs KL change ↓-0.0027 [-0.0042, -0.0013]+0.0006 [-0.0010, +0.0021]-0.0023 [-0.0043, -0.0003]
Pass-through mel-SSIM; without MPO: 0.98350.97980.97700.9770

E3 keeps 78% of the original MPO gain in WILD PQ with only about 28% of the pairs, and improves all three test metrics. E3b also improves WILD PQ, CE, and CU, and gives the largest improvements in test mel-SSIM and log-spectral distance of the three variants; its PANNs KL is unchanged within the interval. So MPO raises WILD PQ even when Audiobox plays no part in ranking the candidates. The listening test above adds a human check that does not rely on the predictor: listeners found PRISM and PRISM without MPO equally good.

We also checked how closely the original preferences follow reference fidelity. In 97.4% of the 10,000 original pairs, the preferred candidate has higher mel-SSIM than the rejected one. Ranking by mel-SSIM alone recovers the preferred candidate in 55.4% of cases and both selected candidates in 27.8%, against chance rates of 20% and 5%; Audiobox alone gives 54.3% and 27.4%. The two signals therefore contribute roughly equally, and the original preferences closely align with fidelity to the reference.

Finally, original MPO improves metrics that were never part of the ranking. On the 1,289 test examples (Section 4.2; Appendix N), log-spectral distance changes by -0.177 dB [-0.211, -0.144], mel-spectral KL by -0.0040 [-0.0062, -0.0018], and PANNs KL by -0.0023 [-0.0043, -0.0003].

AuthorsOpenReview ↗

Response to Reviewer YmPa (2/2)

W1: WILD gains before MPO and after loudness matching

All comparisons below use the same 574 WILD tracks, with every output matched to the loudness of its input:

ComparisonPQ differenceCE differenceCU difference
PRISM minus SonicMaster+0.345 [0.307, 0.385]+0.298 [0.260, 0.335]+0.453 [0.399, 0.506]
PRISM without MPO minus SonicMaster+0.289 [0.252, 0.326]+0.245 [0.210, 0.281]+0.380 [0.329, 0.430]
PRISM minus PRISM without MPO+0.056 [0.051, 0.062]+0.052 [0.047, 0.058]+0.073 [0.066, 0.081]
PRISM minus AudioSR+0.425 [0.376, 0.478]+0.537 [0.482, 0.593]+0.446 [0.377, 0.516]

Two points follow. First, most of PRISM's advantage over SonicMaster is already there before MPO (+0.289 of the +0.345 PQ difference), so it comes mainly from the model trained without any Audiobox-based ranking, and it holds after loudness matching. Second, the way the WILD tracks were selected does not hand large gains to every method: on the same tracks, AudioSR improves PQ over the input by only +0.058 [0.039, 0.078]. The gain is specific to PRISM, not a property of the collection.

Human listeners also clearly prefer PRISM on COMPOUND. It receives 85.0% [68, 97] of choices, against 8.0% for SonicMaster, and 91.4% in their head-to-head comparison. All ten listeners prefer PRISM above chance (sign test, p = 0.001). The new listening test isolates MPO, so on WILD the comparison with SonicMaster rests on the loudness-matched objective results above.

W2 and Q2: Isolating the source of the corpus gain

To separate the factors, we compare PRISM without MPO against SONICMASTER-only, the same architecture trained further on the original SonicMaster corpus. Continued training on the old corpus does not reproduce the gain: SONICMASTER-only improves COMPOUND validation mel-SSIM by only +0.0013. The improvement therefore comes from the new corpus, and the analyses below examine which parts matter. On the 738 random-chain examples from COMPOUND validation and test, where depth and types are drawn independently, the mel-SSIM margins are:

DepthExamplesMargin over SONICMASTER-onlyMargin using only SonicMaster typesExamples using only SonicMaster types
1155+0.078 [0.067, 0.088]+0.068 [0.056, 0.081]95
2143+0.075 [0.065, 0.085]+0.079 [0.064, 0.095]62
3140+0.073 [0.062, 0.084]+0.076 [0.056, 0.095]38
4159+0.072 [0.064, 0.082]+0.081 [0.061, 0.099]29
5141+0.075 [0.063, 0.086]—6

Depth and ordering. The margin is just as large at depth 1, where there is a single degradation, so there is no compounding and no ordering at all. PRISM also restores equally well regardless of the chain order: its mel-SSIM gain over the degraded input is +0.132 [0.125, 0.139] for scenarios that cross production stages and +0.133 [0.126, 0.140] for those in the default order. Both results indicate that ordering does not drive the advantage. These are comparisons across evaluation subsets; a retraining ablation with types, severity, and depth held fixed requires re-rendering the corpus, and we leave it to future work.

New degradation types. The new types add little to the margin. At fixed depth, each additional type absent from SonicMaster changes the margin by +0.003 [-0.004, +0.010], and on examples that contain only SonicMaster's original types the margin is still +0.075 [0.067, 0.083]. PRISM is better even on the degradations SonicMaster was trained for.

Pass-through pairs. The new corpus includes pass-through pairs, which SonicMaster's does not. With it, pass-through mel-SSIM rises from 0.802 to 0.984 (Table 9), so the model learns to leave mastered inputs alone.

Instructions. With an empty instruction, PRISM keeps 95.2% of its gain on compound chains, so the improvement comes from audio modeling, not the new instruction text.

Taken together, the advantage appears at every depth, under both orderings, on the original degradation types alone, and without informative text. No single factor accounts for it, which points to the remaining change—the corpus's degradation parameterization and rendering—as the main source. We discuss this attribution in Appendix Q.

Dear Reviewer,

Thank you for your positive feedback. As the discussion period draws to a close, we would appreciate hearing whether our responses have addressed your concerns or if any further clarification would be helpful.

Best regards,
Authors

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

This work proposes PRISM, a generative model designed for degradation restoration("corrective") and productive mastering ("finishing"). The model is established on SonicMaster with an extra synthetics corpus ("Compound") with a policy optimization stage ("MPO"). The "compound" corpus consists of 24 types of audio effects/distortions, and it has various processing chains with depths varied from 1 to 8. The evaluation shows that in an unseen out-of-domain evaluation set ("WILD"), PRISM improved objective metrics including production quality and mel-SSIM and reached the top performance among the other baselines.

Critical strengths
  • The pipeline that construct the "Compound" corpus is carefully designed. It includes various degradations with proper parameter setup and filtering process to ensure the resulting processing/degradation chain are practical and feasible.

  • Ablation study on where the improvements came from and the probing on model's understanding of identifying degradation deliver convincing support that the model learned proper understanding in tackling "degradations".

Critical weaknesses
  • One concern to the proposed work is the generalizability of proposed system. First, in Table 4, although PRISM shows smaller gap than SonicMaster, but the gap is still large. Second, in the perceptual experiment (Table), PRISM received the same amount of preference as SonicMaster on WILD corpus. These could suggest that PRISM has adapted to tackle in-domain data distribution ("compound test") but not sufficiently capable of tackling out-of-domain data ("WILD").

  • As mentioned in the limitation section, the Audiobox PQ cannot fully resemble true production quality, also the creative choices are not addressed in this work

Questions for authors
  • Is there any specific reason to limit the 30% random scenario with certain order of degradation types, i.e. "mix and mastering → re-recording → transfer and distribution order"? For example, transfer could happen before re-recording.

  • It is interesting to know whether fixed bank of instruction affects model generalizability. Moreover, swapping instruction on proposed corpus do not affect performance much. Could it mean that the model has learned to ignore the instruction and always perform the same task, e.g. always remove degradation and refine the input?

5 comments
AuthorsOpenReview ↗

Response to Reviewer soRj (1/2)

We thank the reviewer for the positive assessment of the corpus design and the probing analysis. Below we respond to generalization, production-quality prediction, degradation ordering, and instruction use. Unless stated otherwise, all intervals are paired 95% bootstrap confidence intervals.

W1: The cross-corpus gap in Table 4

Table 4 asks a specific question: can a linear probe trained on representations from one corpus predict degradation labels on the other? It therefore measures how well the degradation representations transfer, and is separate from restoration quality.

Some gap is expected, because the two corpora implement the "same" degradations differently and have different label base rates. For example, SonicMaster uses measured microphone responses and a fixed 4 kHz low-pass filter, whereas our corresponding degradations are parametric (Table 6). A probe trained on one implementation is being tested on another.

Within that setting, PRISM transfers much better than SonicMaster. Its probes stay above chance for all five degradation groups on the SonicMaster test set, with cAP 0.64. In the reverse direction, SonicMaster's probes reach only cAP 0.38 on our corpus, with stereo at chance. The remaining gap aligns with the implementation differences above.

W1: Generalization to real releases

WILD lies outside both training corpora, and we process every WILD track with a single evaluation instruction that appears in neither the instruction bank nor the training examples. It therefore tests out-of-domain behavior on both the audio and text sides.

On the same 574 tracks, with each output matched to the loudness of its input (ITU-R BS.1770):

ComparisonAudiobox PQ difference, with 95% CI
PRISM minus input+0.483 [0.432, 0.534]
PRISM minus SonicMaster+0.345 [0.307, 0.385]
PRISM minus AudioSR+0.425 [0.376, 0.478]
AudioSR minus input+0.058 [0.039, 0.078]

After loudness matching, PRISM improves predicted PQ on 80.7% of tracks and leads both baselines by a clear margin, so the gains carry over from the synthetic chains to real releases on the objective metric.

On the original five-item listening test, the per-track votes suggest that loudness played a part in the tie. PRISM was preferred on the two tracks where the outputs were within 1 LU of each other. SonicMaster was preferred only on tracks where it was louder (1.7 and 2.5 LU), and even on the track where it was 6.3 LU louder, PRISM matched it in votes.

We also ran a new loudness-matched blind listening test on 15 items (10 from COMPOUND test, 5 from WILD). It isolates MPO. Listeners chose between PRISM and PRISM without MPO, all at -23 LUFS, with trial order and option order shuffled for every listener. Across nine listeners and 135 judgments, PRISM was chosen in 51.1% of cases (95% CI [37.0, 65.2]; chance 50%, p = 0.86). The comparison with SonicMaster on WILD rests on the loudness-matched objective results above.

AuthorsOpenReview ↗

Response to Reviewer soRj (2/2)

W2: Audiobox limitations and creative mastering choices

Audiobox PQ predicts production quality, and as noted in our limitations section, it cannot stand in for every aspect of how a master sounds. We never use it as a differentiable training loss; its influence comes through target selection and the original MPO preference ranking.

To measure how much MPO depends on it, we retrained MPO with preferences set by reference mel-SSIM alone, so Audiobox plays no part in the ranking. Using the 2,778 pairs where mel-SSIM picks the same candidates as the original ranking, we keep 78% of the original MPO gain in loudness-matched WILD PQ: +0.044 [0.040, 0.049]. Re-ranking all 10,000 candidate sets by mel-SSIM alone still improves WILD PQ by +0.025 [0.021, 0.029]. So MPO's PQ gain holds without Audiobox in the ranking. The listening test above shows it comes at no audible cost, and PRISM's main advantage comes from the model trained on our corpus before MPO.

On creative choices: PRISM is designed for corrective restoration and finishing toward the reference-master target described in Section 1 and Appendix A. Within that scope, it respects the choices already present in a finished master, leaving mastered inputs almost untouched (pass-through mel-SSIM 0.978, against 0.821 for SonicMaster). Modeling intentional creative decisions, such as a deliberately lo-fi or heavily compressed sound, is a natural next step that we leave to future work.

Q1: Why constrain the order of random chains?

The ordering constraint applies only to the random-chain component, which makes up 30% of the degraded examples. These chains combine degradation types without an explicit production scenario, so the default order (mixing/mastering → re-recording → transfer/distribution) is a simple way to keep them plausible.

This is a good point: transfer can indeed happen before re-recording, and the corpus already includes this case. The scenario-based component, the remaining 70%, includes chains that cross these stages. In the call-capture scenario, far-end band-limiting (a transfer degradation) comes before device coloration (a re-recording degradation), which is exactly the ordering the reviewer describes. The cassette-transfer and internet-rip scenarios also cross stages.

On the test set, PRISM restores these chains just as well:

Scenario groupExamplesPRISM mel-SSIM gain over inputPRISM margin over SonicMaster
Call capture: transfer before re-recording162+0.169 [0.157, 0.180]+0.150 [0.139, 0.161]
All scenarios crossing stage order466+0.132 [0.125, 0.139]+0.126 [0.121, 0.132]
All scenarios following stage order456+0.133 [0.126, 0.140]+0.127 [0.121, 0.133]

The gains are the same whether or not a scenario follows the default order, and the call-capture case the reviewer raised shows the largest gain of the three groups.

Q2: Does PRISM ignore instructions, and does the fixed instruction bank limit generalization?

PRISM does use the instruction, and how much it relies on it depends on how informative the audio already is:

Evaluation settingEvidence
SonicMaster corpusSwapping instructions removes 39.0% of PRISM's gain.
Compound chainsSwapping instructions removes 1.7% of the gain; an empty instruction retains 95.2%.
Already-mastered inputA mismatched instruction increases 1 − mel-SSIM from 0.022 to 0.055, i.e. the model intervenes more.

On the SonicMaster corpus and on already-mastered inputs, changing the instruction clearly changes the output, so the model is reading the text. On compound chains, the damage is evident from the audio itself, and PRISM repairs the whole chain, named and unnamed faults alike. This is the behavior one wants from a restoration tool given a badly degraded track. Selective control over individual faults within a chain is a natural next step that we leave to future work.

On the instruction bank: each training example is paired with two independently sampled instruction registers (Appendix D), so the model does not see one fixed wording per degradation. It also transfers to wording it has never seen. The WILD evaluation instruction appears nowhere in training, and with it, PRISM improves loudness-matched PQ on 80.7% of tracks.

Dear Reviewer ,

Thank you for your positive assessment and thoughtful feedback. We would appreciate any feedback on whether our responses have addressed your concerns, and would be happy to clarify any remaining points.

Best Regards Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers for their careful reading and helpful feedback. Four points come up across the reviews: (1) perceptual evaluation, (2) dependence on Audiobox, (3) where the corpus gains come from, and (4) how much control the instruction gives. We ran new experiments for each and summarize them here. Changes in the revised paper are in blue. Table numbers below follow the original submission. Intervals are paired 95% bootstrap intervals from 10,000 resamples unless noted.

1. Perceptual evaluation (kGzi, YmPa, soRj). We ran a new blind listening test on the 15 items of the original study (10 from COMPOUND test, 5 from WILD). In each trial, listeners picked the better of PRISM and PRISM without MPO. All clips were set to -23 LUFS, and trial and option order were shuffled per listener. Nine listeners gave 135 judgments and picked PRISM in 69 (51.1%, 95% CI [37.0, 65.2]). Listeners chose the two equally often, so MPO's gains in predicted quality and fidelity come at no audible cost.

On the objective side, the WILD gains hold at matched loudness. Across all 574 tracks, PRISM improves PQ over the input by +0.483 [0.432, 0.534] and leads SonicMaster by +0.345 [0.307, 0.385]. Most of that lead (+0.289) is there before MPO.

2. Dependence on Audiobox (kGzi, YmPa, soRj). We retrained MPO with preferences set by reference mel-SSIM alone, so Audiobox plays no part in the ranking. On the 2,778 pairs where mel-SSIM picks the same candidates as the original ranking, MPO keeps 78% of its WILD PQ gain at matched loudness: +0.044 [0.040, 0.049] against +0.056 [0.051, 0.062]. Re-ranking all 10,000 pairs by mel-SSIM alone still gives +0.025 [0.021, 0.029]. So MPO's gain holds without Audiobox in the ranking, and it comes on top of a lead over SonicMaster that is already there before MPO.

3. Source of the corpus gains (YmPa, soRj). More training on the original corpus does not give the improvement: the same architecture trained further on SonicMaster data gains only +0.0013 mel-SSIM on COMPOUND validation. Against that baseline, PRISM's margin is +0.078 [0.067, 0.088] on single degradations and +0.075 [0.067, 0.083] on examples with only SonicMaster's original degradation types. Its gain is also the same for scenarios that cross the default stage order (+0.132) and those that follow it (+0.133). So the advantage holds at every depth, on the original types, and under either ordering. With the new corpus, which includes pass-through pairs, pass-through mel-SSIM rises from 0.802 to 0.984 (Table 9). A retraining ablation of ordering alone is left to future work.

4. Instruction control (kGzi, soRj). PRISM uses the instruction. On the SonicMaster corpus, swapping instructions removes 39.0% of its gain. On already-mastered inputs, a mismatched instruction raises 1 − mel-SSIM from 0.022 to 0.055. On compound chains, where the damage is clear from the audio, PRISM repairs the whole chain, whether the faults are named or unnamed. Selective control over single faults within a chain is a natural next step for future work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.