acceptodds
Under review as a conference paper at ICLR 2027

One Bit, One Time Slot: Temporally Structured Watermarking for Audio Generation

Abstract

Audio generation models have made high-quality synthetic speech and music easy to produce, and equally easy to copy and redistribute without attribution, raising serious concerns over accountability and copyright. Watermarking offers a practical solution by embedding recoverable payloads into generated audio for ownership verification. Yet, existing efforts either apply the watermark post hoc, leaving it extrinsic to generation and removable by neural denoising or resynthesis, or inject it during generation as an unstructured global perturbation of the latent that entangles all bits and collapses recovery toward chance as the payload grows. In this work, we present TempoMark, a temporally structured generative watermarking framework that treats the generated audio as an ordered trajectory and binds each payload bit to its own temporal footprint. Specifically, TempoMark assigns each bit a dedicated time slot, expands it into chips with a spread-spectrum pattern, and injects the payload as a chip-aligned residual at an intermediate denoising step. Encoding and decoding share this bit-to-slot decomposition, under which each bit is recovered by matched filtering over its assigned chips. On top of this structured backbone, a lightweight neural residual in the decoder is trained to support payload readout, together with a small injection adapter as an acoustic-fidelity correction. Extensive experiments show that TempoMark outperforms state-of-the-art baselines in bit recovery accuracy and imperceptibility, and remains robust against audio perturbations and neural remapping attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Watermarking, Generative audio model, Copyright Protection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_PD9COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TempoMark, a generative audio watermarking method based on temporal structure. The core idea allocates each watermark bit to an independent time slice, spreads information across multiple chips via spread-spectrum coding, injects the watermark during the intermediate denoising stages of diffusion models, and recovers the message using matched filtering, avoiding the inter-bit interference common in traditional generative watermarking. The authors conducted experiments across DiffWave, PriorGrad, and four speech datasets, achieving a watermark recovery accuracy of 98%–100% while maintaining high accuracy under various audio perturbations and neural denoising attacks.

Critical strengths
  • The proposed method is straightforward to understand and easy to implement.

  • Across four datasets and two diffusion models, the bit recovery accuracy reaches 98%–100% while maintaining high robustness under various signal-processing and neural denoising attacks.

  • The authors evaluated the individual contributions of temporal structure, chip alignment, the decoder, and the injection mechanism, showing that temporal structure and matched decoding are the primary drivers of performance improvement.

Critical weaknesses
  • Appendix Table 6 shows that a mere 10 ms time shift drops the WSR from nearly 100% to roughly 50%. This poses a severe practical limitation and undermines the paper's claims of robustness.

  • I suggest that this paper should apply the same temporal spread-spectrum encoding and matched filtering to a post-hoc watermarking baseline and compare it with generative injection, to disentangle whether the performance gains stem from temporal coding or from in-process generative injection.

  • This paper primarily evaluates bit-level WSR rather than full payload recovery or end-to-end provenance verification. Furthermore, the empirical evidence is insufficient to demonstrate robustness against time-shifting, clipping, and adaptive watermark removal attacks.

Questions for authors
  • Appendix Table 6 shows that a about 10 ms time shift drops the watermark recovery accuracy from nearly 100% to roughly 50%. How can the watermark be recovered if the audio is being cropped, spliced, or has unknown temporal offsets? Could this issue be addressed by searching over candidate time shifts?

  • Both spread-spectrum encoding and matched filtering are textbook techniques in classical communications. Compared to traditional spread-spectrum watermarking, what is the technical novelty here? Could you include a traditional post-hoc watermarking baseline using the same encoding and decoding scheme?

  • It is unclear whether the improvements stem from temporal coding or in-process diffusion injection. Can the authors test a post-hoc variant where the same watermark signal is added directly to the generated audio and evaluate it? Such an ablation would directly show whether generative injection is truly needed.

  • The paper mainly reports bit-level WSR, but real-world provenance or copyright verification usually requires recovering the entire payload intact. Could you also report the full message recovery rate? If an attacker knows the watermarking method but not the secret key, can they remove the watermark using targeted audio processing?

5 comments
AuthorsOpenReview ↗

Shift, cropping, and post-hoc comparison

We thank the reviewer for the constructive suggestions. The revision sets out the alignment precondition and the offset estimate in Sec. 5, reports EPR alongside WSR, adds cropping and splicing, and states in the Table 10 caption what it is a negative control for.

Q1&W1: 10 ms shift, cropping or unknown offsets.

Yes, and Discussion and Limitations sets out this precondition with the offset estimate. Despreading applies the chip signs at absolute positions, so a shift beyond one chip applies the wrong signs, which Table 10 is the negative control for, while the operators in Table 5 do not displace the signal. Estimating the lag restores the payload:

shift λfixed originsearchedperfect alignment
±10 ms0.500.941.00
±20 ms0.500.971.00
We further ran cropping and splicing on 64-bit LJSpeech/DiffWave, overwriting or cutting 10% and 50% of the block:
conditionWSRPESQ
clean1.00004.64
one chip / one slot destroyed0.9997 / 0.99183.82 / 3.59
10% overwritten0.95432.64
50% overwritten0.75141.53
10% cut, middle0.72892.67
50% cut, middle0.60861.81
Chip- and slot-level damage is almost free, since spreading carries redundancy inside each bit. Overwriting keeps the grid and costs only the overwritten bits, whereas a cut also displaces the bits after the join.

Q2&W2: Technical novelty and post-hoc baseline.

Spreading and matched filtering are the vehicle, not the contribution. Our finding is that the correspondence between payload bits and time slots governs decoding accuracy, which prior work did not treat as a design variable: structure without alignment sits at chance (0.50), alignment without structure at chance (0.52), and only aligned structure recovers (1.00). The contribution realises that correspondence through chip-aligned injection, a slot partition giving per-bit readout and known erasures, and matched hybrid decoding with the lag estimated. Each element matters:

ablationWSR
full model1.00
w/o CHIP (global remap)0.73
w/o TSS (no structure)0.49
alignment without structure0.52
no lag estimation0.50
lag estimated0.94–0.97
We also built a same-code post-hoc baseline with the same encoder output, slot grid and decoder, so only the injection site differs. With one payload repeated across clips, an attacker estimates the common in-band component CC from the other files and subtracts αC\alpha C, with α\alpha the least-squares coefficient and ρ\rho the fraction of in-band content that CC explains:
variantαρWSR after removalPESQ after removal
post-hoc, +10 dB in-band1.000.9040.6044.38
post-hoc, +20 dB1.000.9900.5254.38
post-hoc, +30 dB1.000.9990.5024.38
in-process (ours)0.780.1330.7271.53
The additive mark is predictable from the other clips, so subtraction drives it to chance and leaves the audio intact, while our carrier gives ρ = 0.13, so the subtraction leaves WSR at 0.73 and costs the audio: a post-hoc mark is shared by every file with the payload, whereas ours is built inside the trajectory.

Q3&W3.1: Temporal coding or in-process injection.

Decoding accuracy comes from the temporal coding, since removing or misaligning the structure drops WSR to chance (Q2). In-process injection is not needed for accuracy, and the same code added to finished audio is no worse under classical distortions or denoisers; it matters under cross-clip collusion, where the additive mark falls to chance and ours keeps 0.73. The w/o INJ row of Table 6 sets injection_mode = latent_init, so it isolates mid-trajectory re-injection rather than in-process placement.

Q4&W3.2: Full payload recovery, and key-unaware removal.

We keep WSR (Eq. 10) and add EPR, the fraction of clips whose whole payload is recovered (released 64-bit checkpoint, 50 clips):

conditionWSREPR
clean0.99940.96
MP3 128 / 64 kbps0.99940.96
codec proxy, 16 kHz + 4 kHz0.99720.82
EPR is 0.96 on clean audio and at both MP3 bitrates while WSR is 0.9994, and under the 16 kHz codec proxy WSR barely moves to 0.9972 while EPR falls to 0.82, i.e. 9 of 50 payloads fail, each by a single bit, which is what verification depends on.
For key-unaware removal the attacker knows the grid and chip length but not the spreading sequence, so it nulls a fixed fraction of every slot:
attackWSR
null 25% of every slot1.000
null 50% of every slot0.998
Nulling half of every slot changes almost nothing, since a null removes only part of a bit's evidence and the surviving chips still add coherently. The stronger key-free attack is the cross-clip subtraction of Q2, which lowers WSR to 0.73 at PESQ 1.53.

Dear Reviewer PD9C:

Thanks for your effort in reviewing our work. As the rebuttal period is nearing the end, please let us know if there are any further concerns or questions. Should there be any outstanding issues or further inquiries, we are eager to address them during the remaining time of the discussion period. If you are satisfied with the clarifications provided in our previous communications, we kindly request you to consider updating your score to reflect the newly added results and discussion, in particular the shift and cropping experiments, the extended payload-recovery results, the ablation against a same-code post-hoc baseline with a cross-clip collusion attack, and the alignment discussion.

Your time and effort are greatly appreciated.

Sincerely,

Authors of Submission 2488

Dear Authors,

I thank the authors for the detailed response. However, I still have several concerns.

(1) Under the independence relation (EPR ≈ WSR^L), the lag-searched WSR of 0.94–0.97 at ±10/20 ms implies an EPR of roughly 2–14% at 64 bits, and a 10% middle cut leaves essentially no clip fully recoverable. The search range, its cost, larger offsets, and its effect on the false-positive rate are not reported.

(2) The key-unaware attack considered (nulling part of each slot) is weak. An attacker who knows the grid can simply desynchronize the signal by small shifts, sample insertions or deletions, or a slight rate change, which is the method’s main weakness and is not evaluated.

Thus, I plan to maintain my original score.

Reviewer_bTSJOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper identifies the research gap that current methods did not embed each bit into the audio distinctly in the audio generation process, thus causing the watermark bits to correlate with the whole audio. Therefore, the paper proposes a more fine-grained method to embed each bit differently in one time slot during the denoising process for diffusion models. With only 2 lightweight networks, it can inject the watermark bits into the diffusion process and decode them correctly. The performance is competitive with AudioSeal both in decoding accuracy and imperceptibility.

Critical strengths
  1. The paper identifies that the bit-to-chunk correlation is important in audio watermarking and designs a preliminary study to find that the correlation is positively correlated with the watermark decoding accuracy. This gap is identified and is ignored by previous work.

  2. The paper uses binary phase-shift keying (BPSK) polarity to encode the watermark and decode it by checking the sign.

  3. This TempoMark is a lightweight framework that is attached to diffusion models. It only needs to train: zadapt and ℓcnnz_{adapt} \text{ and } \ell_{cnn}

Critical weaknesses
  1. The paper title, introduction, and the preliminaries and motivation section are overstated. The method introduced in the paper is only applied to diffusion models. The paper shows effectiveness only in a neural vocoder; it does not prove effectiveness in an autoregressive model or a codec-based generator. I suggest adding diffusion model to the title and toning down your statements in the sections I mentioned.

  2. The statements describing the downsides of post-hoc watermarking and unstructured global perturbation need citations. For example: "Because the watermark is extrinsic to the generative process, it is inherently fragile: neural denoising or resynthesis suffices to strip it away, leaving the content largely unprotected against the attacks it is meant to withstand," "all bits are entangled within one holistic latent modification, offering the decoder no separable, bit-level patterns to exploit," etc.

It is better to provide quantitative results from the citations so the reader can better understand what failure results the paper is describing.

  1. This sentence is repeated in Section 2: "Formally, an audio generator Gθ synthesizes a waveform x = Gθ(c, z) from a condition c and a latent or noise state z." It is defined in the first paragraph of Section 2.

  2. The preliminary experiment is not clearly described. Unstructured, shuffled, and aligned are not explained; neither is the evidence margin. (questions)

  3. In the Method section, not every math notations in the equations are explained. This confuses. Additionally, for your trainable parameters such as z_adapt and l_cnn, it is better to add a superscript to indicate it has parameters.

  4. In your evaluation section, the Tempo watermark underperforms AudioSeal in both WSR and PESQ at 32 bits in all datasets and 64 bits in the LJSpeech dataset.

  5. According to Appendix A.4, the model is highly sensitive to the offset change. It will fail if the watermarked audio is not a multiple of the parameter offset.

  6. The paper describes the 32-bit as the weakest in the wrong-key ROC evaluation on LJSpeech but does not explain why. Intuitively, the higher the bit is the lower the performance is, but the 64-bit is better than the 32-bit.

  7. Timbre and SilentCipher models are not tested here. Timbre is also shown to be very robust, while SilentCipher is good at producing an inaudible watermark. Adding them would strengthen your evaluation.

Questions for authors
  1. "Structural correspondence" and "Unstructured watermarking" need to be explained. What do " structure " and unstructured mean?

  2. What is the superscript cc in TcT_c? What is t\textbf{t}? What is sis_i? There is no definition of sis_i like the m=m1…mi\textbf{m}={m_1\dots m_i}

  3. "which repeats this pulse under known signs so that corrupting part of the slot does not erase the bit." What if corrupting the whole chip?

  4. This also raises a question: if the distortion affects the offset, how does it perform compared to other watermarking methods?

  5. In the threat model paragraph, what does "key" mean in the sentence: "The legitimate owner fully controls the watermarked inference process xw = Gθ(c, z; w), embedding the payload before release and keeping both the key and decoder private."?

6 comments
AuthorsOpenReview ↗

Terminology, notation, and comparison

We thank the reviewer for the detailed review. The revision defines structure and unstructured, adds a notation table and the missing citations, expands the preliminary study, and states the alignment precondition in Sec. 5.

Q1: Structural and unstructured.

A structured watermark writes each bit's chips into its own slot, so the decoder reads bit ii where it was written, whereas an unstructured one rides on a global perturbation and every bit touches every sample. Reading each bit's chip pattern at every slot separates the two, since the aligned profile is diagonal-dominant while the unstructured one is identical up to sign, which is the evidence margin of Fig. 1.

Q2: Notation.

We added a notation table at the end of Sec. 3.1, just before the design subsection, where the cc in TcT_c is a subscript for "chip", not an exponent:

SymbolMeaning
m=(m1,…,mL)m=(m_1,\dots,m_L)payload, mim_i the ii-th bit
si=(si,1,…,si,Nc)s_i=(s_{i,1},\dots,s_{i,N_c})spreading sequence of bit ii, entries ±1\pm1
TcT_cchip duration in samples
NcN_cchips per bit

Q3: Corrupting a whole chip.

A corrupted chip costs almost nothing, since the other chips carry it and the matched branch sums them, while a corrupted slot loses only its bit:

damageWSR
clean1.0000
one chip, 100%0.9997
one slot, 100%0.9918
10 / 25 / 50% of the timeline0.9541 / 0.8719 / 0.7512
Errors stay on the destroyed slots, so WSR falls in proportion and no other bit is affected.

Q4: Offset compared with other watermarking methods.

We ran a new experiment, decoding both baselines and TempoMark under the same shifts:

Shift δ\deltaTempoMarkAudioSealGroot
clean1.0000.990.51
±10 ms0.500.592 / 0.6030.512 / 0.513
±20 ms0.500.504 / 0.4930.513 / 0.515
TempoMark falls to chance at ±10 ms because its grid is absolute (Table 10 is the negative control). AudioSeal keeps 0.60 at ±10 ms but is at chance by ±20 ms, and Groot is at chance even on clean audio. Discussion and Limitations states this precondition and the offset estimate that recovers 0.94–0.97 (Reviewer PD9C, Q1).

Q5: What "key" means.

The key is the set of spreading sequences {si}\{s_i\}, which fix the chip signs the matched filter needs, so the mismatched-sequence condition of Fig. 6 is decoding with a wrong key; the decoder weights are the second private item. The offset τ\tau is not part of the key, since it only has to be shared (Table 11, top).

W1: Title and scope.

Thanks. We narrowed the title to "... Diffusion-Based Audio Generation" and added a scope statement at the end of Sec. 1; see Reviewer cSqc (L2).

W2: Citations for the post-hoc and unstructured claims.

We added the missing citation: the post-hoc claim now cites O'Reilly et al. (2025), who show that neural codecs and denoisers suppress detection of every evaluated post-hoc watermark, and the bit-entanglement claim cites Groot with our own measurement (PD9C, Q2).

W3–W5: Duplicate definition, preliminary study, notation.

We removed the duplicated definition, so that it is now stated once, at its first use in Sec. 2, and made the preliminary study its own subsection with the missing protocol: aligned, shuffled and unstructured differ only in where each bit's evidence is placed, at fixed chip count, energy and decoder. The evidence margin is defined in the text (Q1).

W6: 32-bit WSR and PESQ against AudioSeal.

At 32 bits both reach 1.00 on DiffWave on all four datasets (Table 2), so the WSR gap is confined to PriorGrad, where we reach 0.99–1.00 at 32 bits and 0.98–0.99 at 64 against 1.00, at most 0.02. At 32 bits the fidelity order is the opposite:

32 bit, DiffWaveLJSpeechLibriTTSVCTKLibriSpeech
TempoMark1.711.741.621.75
AudioSeal2.322.191.641.83
A post-hoc mark perturbs a finished waveform and is trained on it, whereas our residual changes the generated content itself, so the same level costs more perceptual quality. At 64 bits the order reverses and we gain 3.24 against 1.64 on LJSpeech/PriorGrad. We removed the blanket imperceptibility claim and report PESQ per setting.

W7: The offset-integer-multiple constraint.

There is no integer-multiple requirement: Table 11 (top) reports 1.00 for τ=0\tau = 0, TcT_c and 50 ms, and the failing rows are mismatched pairs, i.e. a time shift (Sec. 5).

W8: Wrong-key ROC at 32 vs 64 bits.

WSR is per-bit accuracy. Under a wrong key each bit matches at about one half, so the decision concentrates as 1/L1/\sqrt{L} while the owner's score stays near LL, and 32 bits is the hardest setting by construction.

W9: Timbre and SilentCipher.

As suggested we will add both, resampling to a matched rate: SilentCipher has 16 and 44.1 kHz models and Timbre works on the magnitude spectrogram.

Thanks for your clarification. Most of your response is concrete, but can you tell me where the results from Timbre and SilentCipher are? You mentioned it in W9, but I can't find it.

Dear Reviewer bTSJ:

We sincerely appreciate your constructive feedback. As for the results from Timbre and SlentCipher, thank you for pointing this out, and we apologize for the confusion. W9 promised these two baselines for the revision, and these results only became available after submission, which is why they do not appear in the paper yet.

Timbre embeds a message through a CNN that operates on the magnitude spectrogram and reconstructs the waveform with the original phase, repeating the same message along time so that the detector can average its per-frame predictions. SilentCipher embeds a message-conditioned residual in the waveform under an explicit SDR lower bound, which keeps the perturbation inaudible by construction. Both therefore write the payload into coefficients of a finished signal. We re-implemented both designs and trained them on the same 2k LJSpeech clips under the same budget as a 64-bit TempoMark, about 10k optimizer steps:

Payload, budgetTimbreSilentCipherTempoMark
64 bits, ~10k steps0.520.511.00

Under a matched budget, both post-hoc designs remain at the chance level of about 0.5 while TempoMark reaches 1.00, and matching steps is conservative for us, since a TempoMark step also runs the diffusion sampler while a post-hoc step only runs a small encoder. The reason is where the payload lives. In a post-hoc carrier the amplitude of the perturbation is fixed by the fidelity constraint, so every additional bit is paid for by dividing the same energy among more bits, and the per-bit SNR falls roughly as 1/L1/L. TempoMark instead assigns each bit its own temporal slot and lets the block grow with the payload, so adding bits extends the carrier rather than weakening the existing bits, which is why WSR holds at 0.98–1.00 from 8 to 64 bits (Fig. 5, Table 2).

We will add these results and this discussion to the revised version.

Sincerely,

Authors of 2488

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

The paper introduces a method for in-generation audio watermarking as a controlled perturbation to the waveform synthesis model. The perturbation is mapped directly from the payload itself, and results demonstrate good detectability and quality.

Critical strengths
  1. Motivation is principled: embedding the payload bit by bit in an explainable way (pulse waves), combining ML with signal-processing based techniques (BPSK).
  2. Strong detectability with calibrated FPR and good quality, including robustness to codecs and denoising (audio watermarking big weakness).
  3. Reliable up to 64 bits.
  4. Minimal and good quality code.
Critical weaknesses
  1. Figures 3 and 4 indicate that the method is not adaptable to a distortion budget (i.e. attenuating the watermark strength).
  2. Some instances of low quality writing and presentation, just for instance:
    • The paper's WSR (watermark success rate) is ubiquitously mentioned in the literature as "bit accuracy"
    • Table 5 does not need 1-WSR as an additional column.
    • Line 748 is pure slop, tau is effectively an attack parameter (cropping/shifting attack)

Inherent limitations: The method is not at all robust to cropping, and (in its current form) is limited to diffusion generative models.

Questions for authors
  1. What is Figure 6 trying to show? That wrong keys are still strongly detected? Is that a feature or a bug?
  2. Could t_{*} or some other parameter be used to control the watermark strength? (Weakness 1)
  3. Figure 1: Not informative and sloppy, e.g. "evidence margin" ("response at the correct slot minus the average response at others"). Please clarify. As I understand it, it's an ablation of using pulses per bit in each slot (so it should be later in the paper).
  4. Line 241: What are these learnable parameters and why are they necessary?
  5. Design choices:
    • Line 245: Why is alpha learned and not a tuned hyperparameter?
    • Table 4: Why use a codec "proxy" and not the codec itself?
3 comments
AuthorsOpenReview ↗

Figures, strength, and design choices

We thank the reviewer for the detailed feedback. In the revision we have removed the blanket imperceptibility claim from the abstract and state fidelity per setting, stated the embedding level as the strength control in Sec. 3.2 together with a level sweep, defined the evidence margin in the text of Sec. 2 where Fig. 1 is introduced, added a notation table, and narrowed the title and the scope to diffusion-based generation. Below we address Fig. 6, the strength control, the evidence margin, the learnable parameters, the design choices, and the two stated limitations.

Q1: What Fig.6 shows.

Fig.6 shows that a wrong key is not scored like the owner's key, as intended. It keeps the positives of Table 2 and replaces only the negatives, decoding the same watermarked clips with a mismatched spreading sequence, so the audio is identical and only the key differs. At every payload length the curves lie above the diagonal, since a curve on the diagonal would be the defective case.

Q2: Strength control.

Yes, but the parameter is the embedding level, not t∗t_*. The level is the target RMS amplitude of the residual zz (Eq.4), which Sec.3.2 now states. t∗t_* only selects the injection step, so it changes how much of the watermark survives the remaining steps (Fig.8), not how strongly it is embedded. Sweeping the level on the released 64-bit checkpoint:

level (dB)WSRPSNR
−340.95415.92
−260.99515.82
−22 (default)1.00015.60
−101.00011.78
Recovery saturates at the default and a louder watermark only adds distortion, so the default is the operating point of all comparisons.

Q3: Fig.1 and evidence margin.

For bit ii, let ri(j)r_i(j) be the matched-filter response of its chip pattern read at slot jj. The margin is Δi=ri(i)−1L−1∑j≠iri(j),\Delta_i = r_i(i) - \tfrac{1}{L-1}\textstyle\sum_{j\neq i} r_i(j), one value per bit, hence the bit-index axis. It is positive for every bit when the evidence sits where the bit was written, has no consistent sign when the bit-to-slot assignment is shuffled and slot ii holds another bit's evidence, and carries no bit-level information when the perturbation is global. The rows therefore vary only the bit-to-slot correspondence, not the number of pulses per bit, and payloads fail when their evidence is misplaced rather than weak. The component ablation is Table 6.

Q4: Learnable parameters at line 241.

aia_i and bib_i are a per-bit gain and offset on the matched-filter statistic Si(1)−Si(0)S_i^{(1)} - S_i^{(0)} (Eq.7). They turn the raw correlation into a calibrated logit, with bib_i setting the decision threshold of bit ii and aia_i its scale, so that the BCE loss and the mix α\alpha (Eq.8) see the matched branch on the same scale as the CNN branch. They are learnable because that scale and threshold depend on the operating point, unknown before training, and learning them removes a hyperparameter we would otherwise retune. On the 64-bit checkpoint the pair stays at the identity, a=1.0087±0.0078a = 1.0087 \pm 0.0078 and b=5.6×10−4±2.6×10−3b = 5.6\times10^{-4} \pm 2.6\times10^{-3}.

Q5: Design choices.

(a) Learning α\alpha avoids a grid search in which the decoder is retrained per candidate. It is a single scalar (Eq.8) trained with the same BCE loss as the two branches, so it costs one parameter and no extra runs, and it settles near 0.8, so the matched branch dominates the mix. (b) The proxy columns isolate the two basic degradations of lossy coding, band limitation and quantization: codec_proxy_8k resamples to 8 kHz and quantizes, codec_proxy_16k resamples to 16 kHz and low-passes at 4 kHz. We also ran MP3 on the same clips:

conditionWSR
clean0.9994
MP3 128 kbps0.9994
MP3 64 kbps0.9994
codec proxy, 16 kHz + 4 kHz0.9972
codec proxy, 8 kHz + quantization0.9922
MP3 leaves WSR unchanged at both bitrates while both proxies lower it, so the proxies are the harsher test.

W1: Fig.3, Fig.4 vs distortion budget.

Yes. Strength is a single continuous parameter, and its sweep is now in the paper: 12 dB of further attenuation still leaves WSR at 0.954. Figs. 3 and 4 address other axes, a shared PSNR budget and each method's default point, while capacity is Fig. 5, so the strength trade-off comes from the level sweep of Q2.

W2: Writing and presentation.

We have enhanced the writing (also Reviewer bTSJ, Q3).

L1: Cropping.

Removing the start of the clip shifts the remaining slots by ≈93 ms for 10%, which the lag search re-aligns at the same rate as the shift case of Table 10. A 10% head crop therefore leaves the surviving bits at that recovery rate and loses only the bits whose slots were removed, i.e. WSR ≈ 0.86.

L2: Limited to diffusion models.

We focus on diffusion-based generators, DiffWave and PriorGrad on four datasets, because the residual is injected along the sampling trajectory. The title now reads "... Temporally Structured Watermarking for Diffusion-Based Audio Generation".

Thanks for the response. My questions are mostly addressed. I will keep my already positive score.

Dear Reviewer cSqc:

Thank you for the positive assessment and for your careful reading. We are glad the clarifications addressed your questions, and the revision will reflect the changes we described.

Sincerely,

Authors of 2488

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.