acceptodds
Under review as a conference paper at ICLR 2027

CHORD: Learning Watermark Receivers from Unmarked Pairs

Abstract

Watermarks embedded during autoregressive audio generation must support source verification after the generated tokens have been rendered, transformed, and re-encoded. We address this mismatch by learning how the observation channel transforms watermark evidence. For a sender with evaluable likelihood ratios, the observed ratio is the conditional expectation of the source ratio under the unmarked source–observation law. This identity yields a receiver-learning objective and a representation-selection criterion using only paired unmarked data, without waveform reconstruction or key-specific marked training examples. CHORD (CHannel-Observed Response Dictionary) realizes this objective with ordered-candidate sampling, whose rank actions define a small dictionary of local responses shared by all keys and payloads. Learned from unmarked pairs, these responses support composite scoring for joint owner identification and message recovery. On twelve-second music sessions, replacing identity transport with learned responses increases joint owner–payload recovery through two codec compositions from 385 and 379 to 501 and 505 out of 512 sessions, primarily by improving attribution. In speech, unmarked prediction loss selects the receiver, and learned responses yield more gains than regressions under neural codecs, including a held-out operating point. A frozen dictionary transfers to new owner keys without refitting. Together, these results make response learning a practical route to attribution and decoding after audio transformation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Audio Watermarking, AI Safety, Content Provenance

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_LmLuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper claims to model the retokenization mismatch inflicting token-level audio watermarks as a channel degradation, instead of trying to boost token match rate.

Critical strengths

The idea of modeling the retokenization as a lossy channel, and attempting to fit with readily available, unwatermarked audio is novel and principled.

Critical weaknesses

The paper is generally opaque, difficult to read, and presents the results questionably without standard literature metrics. The core mechanism is not explained well. Tables 1 and 2 are uninformative and have no basis in the watermarking literature. Tables 3 and 4 attempt to present TPR and FPR but do so in a peculiar way.

Questions for authors

Please include unwatermarked audio as a quality baseline and FPR check.

2 comments
AuthorsOpenReview ↗

Response to Reviewer LmLu [1/1]

We thank Reviewer LmLu for the assessment that modeling retokenization as a channel and fitting it from unwatermarked audio is novel and principled. We answer the weaknesses in the order of the review, then the question.

W1. Readability and the core mechanism.

A: As our AI Use Statement discloses, we used AI tools to help draft the manuscript and improve the quality of the writing. Every AI-assisted passage then went through a final revision by the authors. We take the readability concern seriously all the same. To make the paper easier to follow, the revision adds algorithm boxes for the sender, the dictionary fitting, and the verifier (Algorithms 1–3, App. F.9), states the endpoints in standard terms (W2 below), and rewrites the Limitations paragraph in plain sentences. The mechanism in one sentence: after any channel, the watermark likelihood ratio equals the conditional expectation of the sender's ratio given the received audio, taken over unwatermarked data, so a receiver can be fit from unwatermarked (generation, received audio) pairs alone.

W2. Standard metrics; Tables 3–4.

A: We did not make the correspondence to standard metrics explicit. The revision does (App. F.9):

Standard metricIn the paper
TPR at 1% FPR (primary metric of Aligned-IS)Table 4
Aligned-IS robustness suite: echo, noise, low-pass, speed, MP3, Opus, DCCRN denoisingTables 16, 29
ASR WER/CERTables 17, 27
Bit accuracy (AudioSeal)Table 20
AUCSec. 5.4, Table 12
FAD (new)App. F.8
Attribution among keys (cf. AudioSeal's message-based attribution)Table 3

Tables 3 and 4 now report rates, and Table 4 is labeled "TPR @ 1% FPR".

W3. Tables 1–2 (joint owner–payload recovery).

A: The task is to identify the generating key among many and to recover its message, which a presence metric cannot measure. Table 11 shows why a joint metric is needed: with identity responses, 299 of 301 failures decode the correct payload under the supplied key, yet the owner is rejected. Detection or bit accuracy alone would hide this. Joint recovery requires both attribution among 10510^5 keys and exact message recovery. The revision explains this and reports Table 1 as rates, next to message accuracy with the supplied key (App. F.9).

Q1. Unwatermarked audio as a quality baseline and an FPR check.

A: Both were in the submission, and we have added stronger versions of each. App. F.9 now lists every table in which unwatermarked audio serves as an FPR check or a quality baseline.

Quality baseline. New: the FAD between marked and unmarked audio matches the FAD between two unmarked sets in every channel. New: in a study with 49 listeners, the watermarked clip was identified 49.7% of the time [46.6, 52.8], and the MOS difference to unmarked audio is +0.013 [−0.083, +0.105] (App. F.8). Already in the submission: Table 17 compares every method with unmarked generations for the same prompts (CLAP; WER/CER for speech).

FPR check. Every cell of Tables 3–4 includes 512 unwatermarked controls. In the new experiments (independent reference keys, nonce search, crop search), no unwatermarked input was accepted. The submission also reports 393,216 owner-absent queries and 4,096 real recordings (Tables 21–22).

On the ethics flag. The research itself, that is, the method, the experiments, and the analysis, was carried out by the authors. Our use of AI tools is disclosed as the ICLR 2027 policy requires, and we take responsibility for all content.

Thanks for the response. I appreciate the effort in disambiguating the paper, and especially the listening study and supplementary materials. Unfortunately I maintain that the presentation is very opaque and low-quality. I will keep my score.

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

This paper proposes CHORD, a receiver-side adaptation method for robust audio watermark detection and message recovery after codec transformations. It learns watermark responses from paired unwatermarked audio and its transformed versions, then reuses the learned response dictionary across keys and messages. Experiments show substantial improvements in joint owner and message recovery for music, although the theoretical guarantees and practical performance under low false-positive rates and more diverse audio settings remain to be further validated.

Critical strengths

The paper presents a clear theoretical motivation and shows substantial improvements in music watermark recovery under codec transformations. The controlled comparisons and broad evaluation provide useful evidence for receiver-side adaptation.

Critical weaknesses
  1. State aggregation and the simplified combination of local responses may discard useful contextual information and ignore dependencies across frames. Their effects on owner ranking and false acceptance rates are not sufficiently analyzed.
  2. The strongest improvements are observed in music experiments, while speech results are less consistent and rely on a relatively small evaluation set. More experiments are needed to establish generalizability.
  3. The receiver assumes access to the nonce, round configuration, and frame alignment. Its robustness under unknown synchronization or realistic deployment conditions remains unclear.
Questions for authors
  1. Can the authors validate the speech results on a larger dataset and report whether the improvements remain significant after correcting for multiple comparisons?
  2. How does CHORD perform when nonce information, round configuration, or precise frame alignment is unavailable? What synchronization assumptions are necessary for practical deployment?
3 comments
AuthorsOpenReview ↗

Response to Reviewer YCYM [1/2]

We thank Reviewer YCYM for the positive comments: (1) a clear theoretical motivation, (2) substantial improvements under codec transformations, and (3) informative controlled comparisons. We answer the three weaknesses and two questions below.

W1. State aggregation and the composite of local responses: effect on false acceptance and owner ranking.

A: Neither can inflate false acceptance, and aggregation helps ranking rather than hurting it.

False acceptance. The bound of the unique-first rule comes from key exchangeability, so it holds whether or not the score models context or across-frame dependence. With independent reference keys the bound applies formally, and no control is accepted in any of the six channels (512 each; App. F.1). The null scores also show no dependence effect: the measured-to-predicted across-key variance ratio is 1.000, the ratio after block shuffling is 1.0005, and the lag-1 autocorrelation is −0.001 (App. F.4). In the submission, 393,216 owner-absent queries gave 2 unique-first events against an expectation of 3.93, and 4,096 real recordings (FMA) gave none.

Owner ranking. We varied the representation directly (8 s joint recovery out of 512; App. F.4):

RepresentationDAC→SNACSNAC→DAC
8 states431429
64 states (used in the paper)435434
512 states429427
2,048 states (no aggregation)363371
Previous-token context added (ctx8)434430

Receivers with 8 to 512 states perform alike. Removing aggregation hurts, which matches the trade-off of Eq. 7: a finer representation keeps more information but costs more estimation error. Adding previous-token context gives no gain, so on this data the discarded context is not what limits the decision.

W2 / Q1. Speech on a larger set, with correction for multiple comparisons.

A: Yes. With 512 sessions instead of 64, all four channels remain significant after Holm correction (same panel and protocol as Tables 2 and 14; App. F.5):

ChannelLearnedSource emissionGainHolm pp
SNAC95.1%90.6%+4.52.3×10−52.3\times10^{-5}
DAC→SNAC86.9%81.6%+5.30.0012
SNAC→DAC89.1%83.0%+6.16.7×10−56.7\times10^{-5}
SNAC 24 kHz (held out)92.0%88.1%+3.90.0012

SNAC 24 kHz was excluded from fitting and selection. Regressions now appear (2–16 sessions per channel), but gains outnumber them 2.7–12.5 to 1. Sec. 5.2 of the revision cites these results.

AuthorsOpenReview ↗

Response to Reviewer YCYM [2/2]

W3 / Q2. Unknown nonce, round configuration, and frame alignment; assumptions for deployment.

A: We now test verification without each of the three. At 12 s the loss is small when the nonce and round order are unknown; the hard case is independently cropped clips under two neural codecs (App. F.2–F.3). Each cell is recovered without the information / recovered with it:

Missing information (12 s)DAC→SNACSNAC→DAC
Nonce and round order, 16 candidate nonces492 / 501498 / 505
Nonce and round order, 256 candidate nonces122 / 126120 / 126
Alignment, session cropped once122 / 126122 / 125
Alignment, each clip cropped independently81 / 11987 / 122

The first row uses 512 sessions and the others 128. Native and DAC lose nothing in any row. At 8 s the nonce search costs more (70–76% of known-nonce recovery with 256 candidates). A single trimmed file corresponds to the session-level row.

No wrong attribution. No unmarked input was accepted in any condition. Under cropping, all 23,484 accepted marked decisions carried the correct payload, so those failures are abstentions. The search scores each key by its maximum over candidates before ranking, which leaves the 1/n1/n bound unchanged.

What deployment requires (App. F.3). (1) The verifier knows a candidate set of nonces, for example the provider's generation log; the round order need not be known. (2) The frame grid can be recovered by search. (3) The observation is long enough: music is stable at 12 s, and at 8 s the search cost becomes noticeable. (4) The time axis is not modified: time stretching and pitch shifting are not handled. Search also multiplies the scoring cost by the number of candidates. Item (4) and per-clip alignment under compositions are stated as limitations.

Processing that keeps the time axis needs no search: the submission already reports full recovery (32/32) under MP3, AAC, and Opus compression, additive noise, low-pass filtering, resampling, and echo (Table 29).

Low false-positive rates and diversity of settings (from the summary).

A: The unique-first decision operates at a nominal 10−510^{-5} level, now formally covered by the independent-key panel (App. F.1). For presence at a 1% FPR, CHORD detects 511–512 of 512 on every channel once the reference term is restored (App. F.6). On diversity, the submission reports marked recovery on 640 source-disjoint sessions and on a new-prompt panel, and false attribution on 4,096 real recordings (Tables 21–22). Each modality uses a single generator; the revision states this as a limitation.

We hope these results resolve the remaining concerns, and we are glad to provide anything further.

I thank the authors for their thoughtful responses and additional experiments, which have satisfactorily addressed my concerns. I therefore maintain my positive recommendation.

Reviewer_4ANoOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CHORD learns receivers for generation-time audio watermarks using paired unmarked generation traces and received audio. It builds on a likelihood-ratio regression identity and constructs a dictionary of local responses shared across owner keys and payloads. The paper evaluates transition-based receiver fitting for music and direct response regression for speech. Its strongest result is a controlled music experiment showing substantial improvements in joint owner acceptance and payload recovery after neural-codec compositions. It also studies unmarked receiver selection, reuse across new keys, and robustness limitations.

Critical strengths
  1. The main music ablation is convincing. The sender, evaluated audio, state representation, keys, and scoring procedure are fixed, so the improvement can be attributed to receiver fitting. Recovery increases from 385 to 501 of 512 sessions on DAC→SNAC and from 379 to 505 on SNAC→DAC. Native-only and count-matched controls largely reproduce these gains.

  2. Learning a reusable response dictionary from unmarked pairs is a useful design choice. The dictionary supports new owner keys without key-specific fitting, and the experiments clearly distinguish owner acceptance from supplied-key payload decoding. This distinction helps explain where the receiver improvements occur.

Critical weaknesses
  1. The connection between the unmarked training objective and the final verification decision remains incomplete.

  2. The practical attribution evidence remains restricted. The main music panel uses owner-derived reference keys and therefore does not inherit the exchangeable-key false-acceptance guarantee. Verification also assumes known public session information and frame correspondence. Speech results require 96 seconds of accumulated observations, while twelve-second native prefixes recover only 4 of 64 sessions. These results support a useful receiver method, but leave uncertainty about reliable verification of short recordings with unknown alignment.

  3. I could not locate audio examples in the manuscript. The quality evaluation uses CLAP similarity and speech recognition error rates, which provide limited evidence about perceptual quality. Please clarify whether marked and unmarked examples, including outputs after codec processing, are available in the supplementary material. Representative samples and a listening evaluation would increase my confidence in quality preservation and could affect my assessment.

Questions for authors

What explains HPT’s failure to identify the generating key even on native audio, despite its strong marked/unmarked discrimination? Can you clarify whether this reflects a limitation of the evaluated configuration for large key panels? For the presence-detection comparison, how does CHORD perform when the omitted reference term is restored?

Are audio samples available? Please provide representative marked/unmarked music and speech examples, with corresponding post-codec outputs, and clarify whether any listening evaluation was conducted.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 4ANo [1/3]

We thank Reviewer 4ANo for the careful review and the positive comments: (1) a convincing main ablation, (2) a useful reusable-dictionary design, and (3) a clear separation of owner acceptance from payload decoding. We answer the weaknesses and questions in order. You noted that audio samples and a listening evaluation could affect your assessment: both are now provided, under W3 in part [2/3].

W1. Link between the unmarked objective and the verification decision.

A: The objective does not affect false acceptance, and it tracks power. A tight theoretical link remains open.

False acceptance. The bound of the unique-first rule comes from key exchangeability and holds for any score function, even a misspecified one. The objective therefore matters only for power.

Power. As the receiver moves from the identity transition to the learned one, the held-out loss falls and recovery rises (joint recovery out of 512; all six interpolation points are in App. F.4):

ReceiverHeld-out unmarked lossDAC→SNAC, 8 sSNAC→DAC, 8 s
Identity2.159239234
Interpolated, t=0.2t=0.21.968422417
Interpolated, t=0.6t=0.61.708435434
Learned (t=1t=1)1.612435434

Loss and normalized owner margin have Spearman −1.0 in all six channel–length cells.

Limits. Across 15 candidates with different representations, the correlation is moderate (−0.30 to −0.53), and Appendix A.5 gives only a sufficient condition for rank preservation. The revision states the tighter link as an open question.

W2. Practical attribution evidence.

A: (a) Owner-derived reference keys. We re-scored the same audio against independent reference keys. Recovery on the compositions is 503/507 (was 501/505), no control is accepted in any of the six channels (512 each), and the bound of Appendix A.6 now applies (App. F.1).

(b) Known session information and frame correspondence. We now verify without the nonce, the round order, and the frame alignment (App. F.2–F.3). Each cell is recovered without the information / recovered with it:

Missing information (12 s)DAC→SNACSNAC→DAC
Nonce and round order, 16 candidate nonces492 / 501498 / 505
Nonce and round order, 256 candidate nonces122 / 126120 / 126
Alignment, session cropped once122 / 126122 / 125
Alignment, each clip cropped independently81 / 11987 / 122

The first row uses 512 sessions and the others 128. Native and DAC lose nothing in any row, and no unmarked input was accepted in any row. Under cropping, every accepted marked decision carried the correct payload, so those failures are abstentions, not wrong attributions. The search scores each key by its maximum over candidates before ranking, so the 1/n1/n bound is unchanged. Independently cropped clips under two codecs remain the hard case; we state this as a limitation.

(c) Short recordings. For speech, the payload is recovered by 24 s and attribution among 100,000 keys by about 48–64 s (native, out of 64; App. F.7):

Duration (s)1216243248648096
Unique first among 100,000 keys48172754636264
Unique first among 1,000 keys1720404562646464
Payload with supplied key5556636364646464

Speech needs longer audio mainly because of its token rate: 12.5 Hz against 50 Hz for music, so the same number of rows takes 4× longer. At an equal row count speech is still lower (54/64 at 48 s; music at 12 s: 512/512), which we attribute to its lower sampling temperature (0.6) and to using only J=8J=8. Attribution among many keys from speech clips of 16 s or less remains out of reach, and we did not test short speech with unknown alignment. Both are stated as limitations.

AuthorsOpenReview ↗

Response to Reviewer 4ANo [2/3]

W3. Audio samples and listening evaluation.

A: The supplementary material now contains four music and four speech examples. Each is given as a marked and an unmarked generation for the same prompt, through all six channels (native and five codec outputs), 96 clips in total. The examples are taken in index order, not selected by outcome. The submission had no listening evaluation. We ran one during the discussion period and added FAD; both find no difference between marked and unmarked audio.

Listening study (App. F.8). The stimuli are the supplementary clips. 50 listeners took part; one was excluded by a low-quality anchor check, leaving 49. In the detection test, a listener hears a marked and an unmarked generation for the same prompt and picks the watermarked one. In the MOS test, marked and unmarked clips are rated from 1 to 5 in random order. A generation-time watermark produces a different sample, not a perturbed copy, so there is no original to compare against; both tests compare the two sets.

Detection accuracy [95% CI]MOS difference, marked − unmarked [95% CI]
Music (600 trials)49.7% [45.6, 53.7]+0.017 [−0.130, +0.173]
Speech (400 trials)49.8% [44.7, 54.8]+0.010 [−0.123, +0.143]
All (1,000 trials)49.7% [46.6, 52.8]+0.013 [−0.083, +0.105]

Listeners were at chance, with an upper confidence limit of 52.8%, and any MOS difference is within about 0.1 on a 5-point scale. Per-channel results (detection 42–56%, MOS difference −0.22 to +0.10) show no consistent direction.

FAD (VGGish; lower is closer). The baseline is the FAD between two halves of the unmarked set. The marked value is the FAD between a marked half and the other unmarked half. Mean and 95% range over 100 random splits:

ChannelUnmarked ↔ unmarkedMarked ↔ unmarked
Native0.165 (0.127–0.223)0.150 (0.113–0.205)
DAC0.162 (0.125–0.223)0.149 (0.117–0.211)
SNAC0.145 (0.107–0.205)0.140 (0.105–0.198)
DAC→SNAC0.139 (0.106–0.195)0.135 (0.104–0.190)
SNAC→DAC0.146 (0.109–0.202)0.144 (0.114–0.213)
NeMo-FSQ0.137 (0.106–0.184)0.137 (0.105–0.184)

In every channel, marked audio is as close to unmarked audio as a second unmarked set is. The two values compare sets of equal size, so the sample-size bias of FAD affects them equally. FAD on CLAP embeddings agrees (App. F.8). This is what output-law preservation (Proposition 2) predicts.

AuthorsOpenReview ↗

Response to Reviewer 4ANo [3/3]

Q1. Why HPT fails to identify the generating key, even on native audio.

A: As you suggest, it is a limitation of the evaluated configuration for large key panels; the presence signal is intact. Diagnostics on native audio at 12 s (App. F.6):

DiagnosticResult
Across-key SD of hit counts: measured / cluster-occupancy model / binomial200.6 / 199.9 / 17.1
Unique-first rate with panel size 10 / 100 / 1,000 / 10,000 / 100,0000.73 / 0.27 / 0.07 / 0.02 / 0.004
Unique first among 100,000 keys at δ = 1 / 2 / 4 (out of 64)Native 0 / 4 / 9; DAC→SNAC 0 / 0 / 0
Signal over across-key SD, 12 s → 96 s1.81 → 2.00 (unique first 0/32 at every length)

HPT assigns green lists at the community level. A recording's hit count for a key then depends largely on how the communities it occupies overlap that key's green communities. Hit counts therefore vary across keys about 12× more than token-level independence (binomial) predicts, and a cluster-occupancy model reproduces the measured spread.

The generating key's signal is only about twice this spread, whereas first place among 100,000 keys needs roughly 4.3 times (Gaussian approximation). Longer audio does not help, because signal and spread grow together. The second row ranks keys by hits standardized by each key's expected green rate, so the result is not an artifact of unequal green rates. This conclusion is about the evaluated setting (h=0h=0, γ=0.25\gamma=0.25, no codec fine-tuning), not about HPT in general.

Q2. Presence detection with the omitted reference term restored.

A: With the term restored, CHORD detects 511–512 of 512 marked sessions on every channel. TP / FP at 12 s, each out of 512, with thresholds calibrated on 256 independent controls at a 1% target FPR:

StatisticNativeDACSNACDAC→SNACSNAC→DACNeMo-FSQ
Raw A(k)A(k) (submission)511 / 1474 / 233 / 113 / 125 / 2116 / 1
With reference term512 / 0512 / 1512 / 3511 / 6512 / 6512 / 2

On SNAC and the compositions, true positives rise from 13–33 to 511–512. Realized false positives are 0–6 of 512 (at most 1.2%), consistent with the 1% target. At the same target, HPT detects 61–123 of 512 (11.9–24.0%) on these channels. Table 4 now reports this statistic, and the sentence "CHORD is not uniformly better on this endpoint" is removed (App. F.6).

Q3. Audio samples and listening evaluation.

A: Please see W3 in part [2/3]: the supplementary material now contains the samples, and we ran a listening study with 49 listeners.

We hope the audio samples, the listening study, and the analyses above address your concerns, and we would be grateful if you would reconsider your assessment in light of them. We are glad to answer further questions.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response: Summary of new results and changes

We thank all reviewers and the AC for their time and careful reading. We are encouraged that the reviewers found the main music ablation convincing and the reusable response dictionary a useful design (4ANo), the theoretical motivation clear and the controlled comparisons informative (YCYM), and the idea of fitting the channel from unwatermarked audio novel and principled (LmLu).

The reviews share four concerns: the false-acceptance guarantee, synchronization assumptions, the link between the training objective and the decision, and the strength of the speech evidence. We ran new experiments on each of these, and on audio quality.

Where to find the new material. To make checking easy for the reviewers and the AC, we collected all new experiments in one place, Appendix F of the revised PDF (pp. 32-39), with every change marked in blue. After the review process, we will move each result to its appropriate section. Each reviewer response is self-contained and includes the tables it needs.

New results at a glance

  1. The guarantee now covers the main panel. We re-scored the same audio against 99,999 reference keys drawn independently of the owner key. Recovery is unchanged (503/507 vs. 501/505 of 512 on the two compositions), and no unmarked control is accepted in any of the six channels (App. F.1).
  2. Verification without the nonce, round order, or alignment. We score each key by its maximum over candidates and only then rank keys. Every key is treated the same way, so the 1/n1/n false-acceptance bound holds for any number of candidates; searching costs power only. With 256 candidate nonces and unknown round order, 12 s recovery on the compositions stays at 95–97% of its known-nonce value (App. F.2–F.3).
  3. Objective vs. decision. Along an interpolation from the identity receiver to the learned one, the held-out unmarked loss and the normalized owner margin are perfectly rank-correlated (Spearman −1.0 in all six cells). Null scores show no measurable across-row dependence (App. F.4).
  4. Speech at 8× the scale. With 512 sessions instead of 64, learned responses improve joint recovery on all four neural-codec channels, significant after Holm correction (pp from 2.3×10−52.3\times10^{-5} to 1.2×10−31.2\times10^{-3}) (App. F.5).
  5. Audio quality. FAD between marked and unmarked audio matches the unmarked-vs-unmarked baseline in every channel. In a new study with 49 listeners, the watermarked clip was identified 49.7% of the time (chance is 50%), and the MOS difference is +0.013 [−0.083, +0.105] (App. F.8). The supplementary material contains 96 audio clips: four music and four speech examples, marked and unmarked, through all six channels.
  6. Presence with the reference term restored. True positives at a 1% target FPR rise from 13–33 to 511–512 of 512 on SNAC and the two compositions (App. F.6).
  7. Why HPT does not identify keys in a large panel. Its score spread across keys is about 12× what token-level independence predicts (App. F.6).

Changes in the revised PDF (all in blue)

  • Appendix F (new): F.1 independent reference keys; F.2 unknown nonce and round order; F.3 unknown alignment and deployment requirements; F.4 loss vs. decision; F.5 speech with 512 sessions; F.6 presence and HPT diagnostics; F.7 speech duration; F.8 audio samples, FAD, and the listening study; F.9 algorithm boxes and a map to standard metrics.
  • Main text, only where a statement changed: Tables 3–4 report rates, and Table 4 is TPR @ 1% FPR with the reference-term statistic; Secs. 5.2, 5.3, and 5.5 cite the new speech, key-panel, and quality results; the Limitations paragraph is rewritten.
  • Ethics Statement: a note on the listening study.
Area_Chair_FNG2OpenReview ↗

ICLR Review Discussion – Follow-up on Author Rebuttal

Dear Reviewers,

Thank you for your time and effort in reviewing this submission.

The authors have submitted their rebuttal. Could you please take a moment to review their responses and assess whether your concerns have been adequately addressed? If you have any remaining concerns or new questions, please feel free to raise them in the discussion.

If you have already responded to the authors, please kindly disregard this reminder.

Thank you again for your valuable contributions to the review process.

Best,

AC

Dear Area Chair, dear fellow Reviewers, dear Authors.

I apologize for my previous rushed response to the rebuttal. I do appreciate the listening study and supplementary materials. My big concern is not whether the presented method "works", but more about the quality of the paper, which in my opinion is not ready for publication at ICLR. The writing was unclear throughout. Here are indicative examples to support my assessment of lack of clarity and low quality:

  • Line 129: what are "public session coordinates"?
  • Line 289: "Music uses a fixed 64-state representation; speech selects state count and α using held-out unmarked loss.", why different treatment for each audio paradigm? Seems bloated.
  • Line 317: "Scoring reuses two symbol contributions per row and key across all payloads; Appendix B.6 times 100,000 keys and sixteen payloads."
  • Section 5, the entire intro (lines 323-370) is convoluted, in particular: the authors mention using MusicGen to generate "sessions of three separate observations", which as I understand it, translates to "audios", "carrying the same 4bit payload" --> so are they watermarked using some model? And only 4bits? Also: "Joint recovery requires generating-key acceptance and exact payload recovery. Supplied-key payload recovery evaluates decoding without owner acceptance. The music panel contains K0 and 99,999 owner-derived HMAC decoys; results are empirical fixed-panel outcomes."
  • Line 411: "Entries are indexed by action and state, not key, so reuse holds by construction;", what "reuse" holds?
  • Line 461: "On native audio, [HPT]'s median conservative rank is 5,571, with no generating-key tie for first.", what is a median conservative rank and a generating-key tie?

The paper's experimental task is very very simple: showcase that learning the channel's response (abstractly shown in Fig. 1(b), the "learned receiver", and explained in Section 4) increases payload m recovery for watermarked audio, per channel and per watermarking sender/receiver pair. I look at Tables 1, 2 and do not see the takeaway. The corrections to Tables 3 and 4 are indeed improvements, and Table 4 is more informative now. I would kindly ask the rest of the Reviewers and Area Chair to verify this themselves and discount my review if the main takeaway is obvious to them.

Again, I apologize for the harsh criticism. The authors did disclose their AI usage so the flag was unnecessary, and all my above concerns fit well into the "Critical Weaknesses" section of the review. Thank you for prompting me to clarify!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.