acceptodds
Under review as a conference paper at ICLR 2027

Who Is AI Singing As? Singer-Timbre Misuse Detection in AI Music

Abstract

AI-generated music can imitate the vocal identity of real singers, raising a question beyond DeepFake detection: who is AI singing as? We study this problem as real–synthetic singer verification and identify a substantial identity gap: pretrained speaker-verification models that perform well on real singing can fail severely on synthetic vocals. To study this gap, we develop TimbreTrace, which adapts pretrained speaker-verification models using mixed real–synthetic supervision. Our analysis shows that pretrained models retain source-singer characteristics even when converted vocals are perceived as the intended target singer, while post-training shifts representations toward the target singer and improves real–synthetic verification. We further evaluate how these gains transfer to Zero-day AI Songs. These results establish the real–synthetic singer-identity gap as a challenge for singer verification and show that real–synthetic supervision can markedly reduce it.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI Music, Copyright Protection, Singer Verification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_VkxDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the TIMBRETRACE system that addresses the problem of identifying whether an AI generated singing voice is similar to the target singer. It adapts pretrained speaker verification models to real and synthetically generated singing data. Experiments show substantial improvements on a controlled benchmark and smaller improvements on an external collection covering ten unseen singers. The paper also investigates whether pretrained models retain source-singer identity despite humans perceiving the intended target.

Critical strengths
  1. The main contribution of this paper is to address the problem of determining whose voice AI-generated music resembles, not just whether the music was generated by AI. This is an interesting problem that is practical as musicians want to know whether their vocal identity was imitated by AI.

  2. The results show that this paper substantially improves the pretrained models in terms of Equal Error Rate: from 45.28% to 9.08% using ERes2Net in real–synthetic experiments. This paper also shows improvements over real-only adaptation and an AAM-Softmax baseline.

Critical weaknesses
  1. The explanation in the source–target setting is not clear. Section 4.2 states that pretrained models frequently favor the source singer, while Figure 2 shows higher average target similarity for all three pretrained models; these findings could be consistent, but we still do not have the proportions at trial level. In the human evaluation, the models’ disagreement with the listeners on the same examples is not reported, so the main explanation remains weak.

  2. The evaluation on the external datasets does not isolate the benefit of the synthetic supervision. Table.2 compares full-adapted models vs. "pretrained models", while Table.4 ablates on the training datasets based on the pooled benchmark scores. The authors’ constructed benchmark shares singer identities and conversion checkpoints between training and testing; an evaluation of the “real-only” models externally is thus needed to distinguish the benefit of synthetic supervision from adaptation to singing.

Questions for authors
  1. What proportion of examples favor the source before and after adaptation, particularly among the human-validated songs? Do source references use different recordings from the conversion inputs?

  2. Can the existing real-only and full-adaptation models be compared separately on controlled real–synthetic trials and the external collection? Are all verification trials in Table 2 real–synthetic pairs, or does the evaluation also include real–real pairs?

2 comments
AuthorsOpenReview ↗

Response to W1 & Q1

We thank the reviewer for requesting clearer evidence about source–target preference and comparisons that isolate the benefit of synthetic supervision. We have added trial-level statistics, a larger sample of model judgments complementing the human validation, and separate controlled and external real–synthetic evaluations.

W1 & Q1. Source–target preference before and after adaptation

We agree that average similarities alone do not establish how frequently individual examples favor the source singer. We therefore report source-preference proportions over 11,984 segment-level source–target comparison trials.

For each trial, we compare a synthetic segment's mean cosine similarity to source-singer references with its mean cosine similarity to target-singer references. A trial is source-favored when the former exceeds the latter. Each segment contributes one comparison trial; 11,984 is the number of trials, not the number of songs. We use the same trials and references before and after adaptation.

BackboneSource-favored before adaptationSource-favored after adaptation
ERes2Net1,611/11,984 (13.44%)54/11,984 (0.45%)
CAM++1,695/11,984 (14.14%)438/11,984 (3.65%)
ECAPA-TDNN1,311/11,984 (10.94%)20/11,984 (0.17%)

The corresponding mean cosine similarities are:

BackboneSettingSource similarityTarget similarity
ERes2NetPretrained0.38940.4885
ERes2NetTIMBRETRACE0.07180.5599
CAM++Pretrained0.32470.4312
CAM++TIMBRETRACE0.11010.4380
ECAPA-TDNNPretrained0.23910.3583
ECAPA-TDNNTIMBRETRACE0.05370.5766

These results clarify why higher average target similarity can coexist with source-favored individual trials. Before adaptation, 10.94%–14.14% of the analyzed trials favor the source, despite higher target similarity on average. After adaptation, this proportion decreases to 0.17%–3.65% across the three backbones. Meanwhile, mean source similarity decreases and mean target similarity increases, widening the target–source similarity gap.

Human validation and model judgments

The original human validation covered 50 songs with three listeners. We extended it to 500 newly sampled segments, each judged by 3 blinded listeners as closer to the source or the target singer. Listeners judged 486/500 (97.2%) of the synthetic samples as closer to the intended target singer.

On the same 500 samples, we compare each model's source–target preference (defined above) with the listener majority:

BackboneAgreement before adaptationAgreement after adaptation
ERes2Net83%98%
CAM++82%96%
ECAPA-TDNN85%96%

Wording and reference selection

We will replace the imprecise wording "frequently favor the source" with these measured proportions. These results characterize the change in source–target preference on the analyzed samples; they do not by themselves establish that this change explains all verification improvements.

Regarding reference selection, all source-singer references come from recordings different from the conversion input. We exclude the entire recording used as the conversion input, not merely the corresponding segment.

AuthorsOpenReview ↗

Response to W2 & Q2

W2 & Q2. Real-only versus Full adaptation on controlled and external RS trials

We agree that comparisons against pretrained models alone do not isolate the benefit of synthetic supervision beyond adaptation to real singing. We therefore compared Real-only and Full adaptation separately on controlled RS trials and external zero-day RS trials, using models retrained under the revised singer-disjoint protocol described in our general response.

Both evaluations below contain RS pairs only, with a real reference and a synthetic candidate. EER and TPR@1%FAR are reported in percentages.

BackboneTraining settingControlled RS EER ↓Controlled RS AUC ↑Controlled RS TPR@1%FAR ↑Zero-day RS EER ↓Zero-day RS AUC ↑Zero-day RS TPR@1%FAR ↑
ERes2NetReal-only14.290.935536.8314.150.941045.12
ERes2NetFull11.960.947643.2913.570.939944.69
CAM++Real-only15.710.920832.7914.080.937640.62
CAM++Full11.000.955050.5013.040.946645.98
ECAPA-TDNNReal-only13.670.935741.3315.330.918638.26
ECAPA-TDNNFull11.880.950644.6714.820.929544.27

On the controlled RS subset, Full adaptation improves over Real-only adaptation across all three metrics for every backbone. These comparisons support a benefit from including synthetic supervision beyond adaptation to real singing under the revised protocol.

On the external zero-day RS subset, Full adaptation achieves lower EER for all three backbones and improves all three metrics for CAM++ and ECAPA-TDNN. For ERes2Net, EER decreases from 14.15% to 13.57%, while AUC and TPR@1%FAR are slightly lower. Thus, the external benefits depend on the backbone and metric rather than being uniform.

Because these evaluations contain only RS pairs, the observed improvements cannot be attributed to better performance on RR trials. We will include these comparisons and their qualifications in the revised manuscript.

Q2. Composition of the zero-day evaluation in Table 2

The overall zero-day results in Table 2 include both RR and RS pairs, rather than exclusively RS pairs. The evaluation contains 29,856 verification pairs:

Pair typePositive pairsNegative pairsTotal
Real–real (RR)7,4647,46414,928
Real–synthetic (RS)7,4647,46414,928
Total14,92814,92829,856

Positive pairs share the same singer identity, whereas negative pairs have different singer identities. The reference is always a real recording; the candidate is real for RR pairs and synthetic for RS pairs.

All 29,856 pairs contribute to the pooled evaluation in Table 2. The separate zero-day RS results above use only the 14,928 RS pairs. These verification pairs are distinct from the source–target comparison trials reported in W1 & Q1. We will clarify the evaluation procedures and their respective counts in the revised manuscript.

Reviewer_TQ72OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the real-synthetic singer verification gap, where speaker / singer verification systems can perform badly on verifying synthetic song voices to the singer whose voice is being imitated. This is a challenge with real-world implications because it can allow synthetic songs to not register as the person whose voice is being imitated, which may not trigger the appropriate copyright / attribution / personality-right systems related to the singer being imitated. The authors argue that pre-trained models can retain source singer characteristics even after target singer conversion. The authors propose a benchmark of 22 singers, along with a triplet-loss based fine-tuning method to improve representations toward the target singer, which improves real-real verification and verifying synthetic songs to the target singer (on their benchmark and on AI songs from unseen singers)

Critical strengths

The paper claims that mixed real-synthetic supervision can address a gap in the existing verification models and their ablation studies (specially Table 4) demonstrate that using Real-Synthetic (RS) pairs for supervision provides performance boost over the baseline and Real-only supervision. The performance of RS supervision over Real-only supervision offers some support to one of the central claims of the paper, which is why this paper deserves consideration.

Critical weaknesses

As the authors note, the benchmark shares the same singers. One checkpoint per singer might mean that they have used the same checkpoints in training and evaluation, so it is not fully established that the proposed method is actually generalizing instead of learning checkpoint-specific artifacts. While the unseen singer performance on Table 2 does support the generalization claim, it is unclear how much of these pairs are real-synthetic and how much of the pairs are real-real. While at least some of the gap in performance between Table 2 and Table 3 can be attributed to out-of-distribution data in Table 2, it is important to clarify the concerns of the generalization, since it is one of the central claims of the paper.

Questions for authors

Clarifications:

  1. Can the authors confirm whether each checkpoint was unique to a singer, and the same checkpoint is used across train and eval sets?
  2. Can the authors use the checkpoints from Table 4 and run them on the zero-day sets? Specifically, the RS pair performance there can provide useful insight about generalization.
  3. Can the authors provide further details about the zero-day set data distribution (how many pairs were real-real and how many pairs were real-synthetic)?
1 comment

We thank the reviewer. We retrained models with globally singer-disjoint splits and added pooled and RS evaluations.

W1 & Q1. Singer/checkpoint overlap and generalization

Yes, each singer has one singer-specific conversion checkpoint. The original benchmark shared singers and checkpoints across training and evaluation while holding out source recordings; it therefore did not establish unseen-singer/checkpoint generalization.

The revised benchmark contains 44 distinct singers and 44 checkpoints in two non-overlapping groups of 22: the original SVC group (so-vits-svc 4.0 and DDSP-SVC) and the new RVC group. Each group is split into 14/4/4 training/validation/test singers, yielding 28/8/8 globally disjoint singers and checkpoints. Synthesis occurs exclusively within splits, with both source and target singers belonging to the same split.

We regenerated part of the synthetic collection, resampled triplets, and retrained all three backbones and AAM-Softmax baselines. As detailed in our general response, TIMBRETRACE reduces controlled RS EER from 21.42% to 11.96%, 19.50% to 11.00%, and 16.46% to 11.88% for ERes2Net, CAM++, and ECAPA-TDNN, respectively.

These gains persist for held-out singers and checkpoints. Generator models remain represented across splits, so this is not a held-out-generator-family evaluation.

W1 & Q2. Table 4 training settings on external and controlled RS trials

We evaluated Real-only (R), Synthetic-containing (S), and Full (F) using models retrained under the revised singer-disjoint protocol, rather than the original Table 4 checkpoints. The same models were evaluated on the external zero-day collection and controlled Own Test split.

EER and TPR@1%FAR (TPR) are percentages; lower EER and higher AUC/TPR are better. Bold marks the best value within each backbone and evaluation subset, including ties.

External zero-day collection

Pooled results use all 29,856 RR+RS trials. RS-only results use the 14,928 RS trials.

BackboneSettingPooled EERPooled AUCPooled TPRRS EERRS AUCRS TPR
ERes2NetR13.340.945550.6414.150.941045.12
ERes2NetS13.770.935948.6614.040.930445.12
ERes2NetF13.400.941749.2513.570.939944.69
CAM++R13.800.942446.4614.080.937640.62
CAM++S12.600.950749.5712.880.949647.37
CAM++F13.140.947247.3713.040.946645.98
ECAPA-TDNNR14.480.929143.6215.330.918638.26
ECAPA-TDNNS13.610.939248.1813.930.934645.77
ECAPA-TDNNF14.310.934447.7014.820.929544.27

For CAM++ and ECAPA-TDNN, both settings with synthetic data (S and F) improve all external RS metrics over Real-only; S performs best in both pooled and RS evaluations.

For ERes2Net, Full improves RS EER (14.15% → 13.57%) but slightly lowers AUC (0.9410 → 0.9399) and TPR (45.12% → 44.69%). Real-only performs best on pooled metrics. External benefits therefore depend on backbone and metric.

Controlled benchmark (Own Test)

Results below use RS trials only:

BackboneSettingRS EERRS AUCRS TPR
ERes2NetR14.290.935536.83
ERes2NetS13.080.937438.62
ERes2NetF11.960.947643.29
CAM++R15.710.920832.79
CAM++S12.460.947949.08
CAM++F11.000.955050.50
ECAPA-TDNNR13.670.935741.33
ECAPA-TDNNS12.080.946345.17
ECAPA-TDNNF11.880.950644.67

Both settings with synthetic data (S and F) improve all controlled RS metrics over Real-only for every backbone. Full achieves the best EER/AUC throughout; S achieves the best ECAPA-TDNN TPR.

These results support benefits from synthetic supervision beyond real-singing adaptation on the controlled benchmark, with varying external gains. RS-only improvements cannot be attributed to better RR performance. We will include these results and qualifications in the manuscript.

W1 & Q3. Zero-day composition and RS-only reporting

Table 2 pools 29,856 trials, equally divided between real–real (RR) and RS:

Pair typePositiveNegativeTotal
RR7,4647,46414,928
RS7,4647,46414,928
Total14,92814,92829,856

Positive pairs share a singer identity; negative pairs do not. Synthetic identity refers to the intended target singer. References are always real, candidates are real in RR and synthetic in RS.

All trials contribute to Table 2's pooled results; the separate RS evaluation uses only the 14,928 RS trials. We will clarify this composition and report RS-only results in the manuscript.

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

The paper asks a simple question: when an AI song copies a singer’s voice, can we tell whose voice it is copying? The paper finds that existing speaker-verification models get this wrong on AI vocals, so they fine-tune those models on real and AI singing pairs (TIMBRETRACE) and test whether this approach still works on AI songs from singers the model has never heard.

Critical strengths

The problem framing and the source-versus-target analysis in Figure 2 are a real contribution. Other than this, I could not find much of strength in this submission.

Critical weaknesses

Major weakness: The benchmark is small: 22 singers, 110 recordings, two conversion systems from 2023. No modern SVS or full-song generators (RVC, Seed-VC, Suno-style systems). Zero-day has 94 AI recordings from 10 singers. Further, every AI sample’s source singer is one of the 22 reference singers. What fraction of RS negatives have source singer equal to the reference? More so, singer identities are shared across benchmark splits (Section 4.1, “this setting evaluates transfer to held-out source recordings rather than unseen singers”). So the 45% to 9% EER drop in the abstract and Table 3 is on 22 seen singers and 22 seen checkpoints, and the model may be learning checkpoint features. The unseen-singer result in Table 2 - 18.7% to 16.7% EER - is on a very small set.

Other comments:

  1. References are incomplete and probably there are errors as well. For example, where Sahidullah et al is published is missing. There are several papers from arxiv but many of them are already published - the authors should use published references.

  2. Zero-day TPR@1%FAR is 34% to 40% is very low for a misuse-detection application (the system catches only a third of imitations at a 1% false-accept rate)? Did the authors try synthetic songs available in online social media - not only at YouTube?

  3. Human validation uses 50 songs and three listeners, one of them a co-author, with no agreement statistic. This is a very weak analysis.

  4. The only trained baseline is AAM-Softmax. The authors should use fine-tuned SOTA deepfake detectors in baselining the results.

Questions for authors

Major weakness: The benchmark is small: 22 singers, 110 recordings, two conversion systems from 2023. No modern SVS or full-song generators (RVC, Seed-VC, Suno-style systems). Zero-day has 94 AI recordings from 10 singers. Further, every AI sample’s source singer is one of the 22 reference singers. What fraction of RS negatives have source singer equal to the reference? More so, singer identities are shared across benchmark splits (Section 4.1, “this setting evaluates transfer to held-out source recordings rather than unseen singers”). So the 45% to 9% EER drop in the abstract and Table 3 is on 22 seen singers and 22 seen checkpoints, and the model may be learning checkpoint features. The unseen-singer result in Table 2 - 18.7% to 16.7% EER - is on a very small set.

Other comments:

  1. References are incomplete and probably there are errors as well. For example, where Sahidullah et al is published is missing. There are several papers from arxiv but many of them are already published - the authors should use published references.

  2. Zero-day TPR@1%FAR is 34% to 40% is very low for a misuse-detection application (the system catches only a third of imitations at a 1% false-accept rate)? Did the authors try synthetic songs available in online social media - not only at YouTube?

  3. Human validation uses 50 songs and three listeners, one of them a co-author, with no agreement statistic. This is a very weak analysis.

  4. The only trained baseline is AAM-Softmax. The authors should use fine-tuned SOTA deepfake detectors or spoofing-aware speaker verification in baselining the results.

Post rebuttal: I am updating the score to 2 (from 1). Reason is mentioned in my reply to authors response.

1 comment

We thank the reviewer for the constructive feedback. We expanded the benchmark, introduced globally singer-disjoint splits, and retrained and evaluated all three backbones, including AAM-Softmax baselines.

W1 & Q1. Benchmark coverage, singer/checkpoint overlap, and RS negatives

We agree that the original benchmark shared singers and conversion checkpoints across training and evaluation and therefore did not establish generalization to unseen singers or checkpoints.

The revised benchmark contains 44 distinct singers and 44 singer-specific checkpoints across so-vits-svc 4.0, DDSP-SVC, and the newly added RVC. It comprises two non-overlapping groups of 22 singers: the original SVC group (so-vits-svc 4.0 and DDSP-SVC) and the RVC group.

Each group is split into 14/4/4 training/validation/test singers, yielding 28/8/8 globally disjoint singers and checkpoints. Synthesis occurs exclusively within each split: both source and target singers belong to that split.

We regenerated part of the synthetic collection, resampled training triplets, and retrained all three backbones. Controlled real–synthetic (RS) EERs (%) are:

BackbonePretrained ↓TIMBRETRACE ↓
ERes2Net21.4211.96
CAM++19.5011.00
ECAPA-TDNN16.4611.88

Improvements persist with evaluation singers and checkpoints excluded from training. Since construction and trials changed, these scores should be compared within the revised protocol, not directly with original scores.

The test split contains eight singers. Generator models remain represented across splits, so these results support held-out-singer/checkpoint generalization, not held-out-generator-family generalization. Modern singing synthesis and full-song generation remain underrepresented.

In the rebuilt Own Test split, 777/2,400 RS negatives (32.38%) have a synthetic candidate whose source singer matches the reference. They are negative because the intended target differs from the reference singer, testing whether residual source cues cause false accepts.

W2 & Q2. References

Thank you for identifying the incomplete reference Sahidullah et al. We will complete it, check other incomplete references, and replace arXiv citations with published versions where available.

W3 & Q3. External evaluation at 1% FAR and platform coverage

We agree that performance at a strict false-accept rate is important for assessing practical limitations. We evaluated the retrained models on the external zero-day collection of 94 AI recordings covering 10 singers. Pooled results include real–real (RR) and RS trials. All values below are TPR@1%FAR (%):

BackbonePretrained pooled ↑TIMBRETRACE pooled ↑TIMBRETRACE RS-only ↑
ERes2Net26.5849.2544.69
CAM++37.0847.3745.98
ECAPA-TDNN28.6747.7044.27

All three models improve pooled TPR over their pretrained counterparts, including on RS-only trials.

These scores measure singer-identity verification. Establishing misuse additionally requires evidence of AI generation and authorization. The collection comes exclusively from YouTube; other social-media platforms have not been evaluated, and we will clarify this limitation in the manuscript.

W4 & Q4. Human validation

We agree that 50 songs and three listeners, including one co-author, provide limited perceptual evidence. The 48/50 majority-vote result measures agreement with intended target labels; inter-rater agreement is a separate quantity and was not reported in the original manuscript.

We plan a blinded follow-up with five independent non-co-author listeners, covering both SVC and RVC groups and reporting inter-rater agreement and majority-vote results. This follow-up remains pending and is not evidence for the current results.

W5 & Q5. Task-matched baselines

Thank you for suggesting additional baselines. TIMBRETRACE verifies whether a real or synthetic candidate matches a real reference singer. A converted vocal is a positive trial when its intended target matches the reference, whereas deepfake detection labels it synthetic and standard spoofing-aware verification ordinarily treats it as a spoof. Directly using those scores would therefore evaluate a different target; these architectures would require adapted objectives and scoring.

We reran AAM-Softmax on the same three backbones under the revised protocol. Controlled RS EERs (%) are:

BackboneAAM-Softmax ↓TIMBRETRACE ↓
ERes2Net13.5411.96
CAM++15.4611.00
ECAPA-TDNN13.2111.88

TIMBRETRACE improves over this task-matched baseline for every backbone. These results do not establish superiority over deepfake-detection or spoofing-aware architectures adapted to our task.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their constructive feedback. A shared concern is that overlapping singers and singer-specific conversion checkpoints in the original benchmark made generalization difficult to distinguish from memorization. We therefore expanded the benchmark, introduced globally singer-disjoint splits, and retrained and evaluated all three backbones, including AAM-Softmax baselines. The updated results support our main conclusions. We summarize the changes below and address specific concerns in our individual responses.

1. Expanded benchmark and globally singer-disjoint evaluation

The revised benchmark contains 44 distinct singers and 44 singer-specific conversion checkpoints across so-vits-svc 4.0, DDSP-SVC, and RVC. It comprises two non-overlapping groups of 22 singers: the original SVC group (so-vits-svc 4.0 and DDSP-SVC) and the newly added RVC group.

Each group is partitioned into 14/4/4 training/validation/test singers, yielding 28/8/8 globally disjoint singers and checkpoints overall. Cross-singer synthesis occurs exclusively within each split, so both source and target identities are disjoint across splits.

We regenerated part of the synthetic collection, resampled training triplets within splits, and retrained ERes2Net, CAM++, and ECAPA-TDNN from their pretrained models using TIMBRETRACE and AAM-Softmax. This protocol evaluates held-out singers and singer-specific checkpoints, but not held-out generator families, since generator models are represented across splits.

2. Updated benchmark results, including AAM-Softmax

The table compares pretrained models, AAM-Softmax, and TIMBRETRACE under the revised protocol. RR and RS denote real–real and real–synthetic verification. Pooled results combine our benchmark's RR and RS subsets with OpenSinger and M4Singer RR subsets. EER is in percent.

BackboneAdaptationRR EER ↓RS EER ↓RS AUC ↑Pooled EER ↓Pooled AUC ↑
ERes2NetPretrained16.3321.420.870016.070.9194
ERes2NetAAM-Softmax10.0013.540.93749.890.9651
ERes2NetTIMBRETRACE9.6211.960.94769.410.9667
CAM++Pretrained16.3819.500.878117.720.9073
CAM++AAM-Softmax12.5815.460.921512.250.9477
CAM++TIMBRETRACE10.5811.000.955010.210.9588
ECAPA-TDNNPretrained12.5816.460.915214.310.9288
ECAPA-TDNNAAM-Softmax11.5413.210.941510.580.9578
ECAPA-TDNNTIMBRETRACE10.6711.880.950610.270.9610

TIMBRETRACE reduces RS EER from 21.42% to 11.96%, 19.50% to 11.00%, and 16.46% to 11.88% for ERes2Net, CAM++, and ECAPA-TDNN, respectively. It also outperforms AAM-Softmax in RS EER/AUC and pooled EER/AUC for every backbone.

These gains persist with evaluation singers and checkpoints excluded from training. Because construction and evaluation trials changed, comparisons should be made within the revised protocol, not directly against original benchmark scores.

3. Updated external zero-day results

We evaluated the three retrained TIMBRETRACE models on the external zero-day collection. Results pool RR and RS trials; EER and TPR@1%FAR are percentages.

BackboneAdaptationEER ↓AUC ↑TPR@1%FAR ↑
ERes2NetPretrained18.700.889826.58
ERes2NetTIMBRETRACE13.400.941749.25
CAM++Pretrained18.840.895937.08
CAM++TIMBRETRACE13.140.947247.37
ECAPA-TDNNPretrained20.850.866628.67
ECAPA-TDNNTIMBRETRACE14.310.934447.70

All three models improve over their pretrained counterparts across EER, AUC, and TPR@1%FAR, supporting transfer beyond the controlled benchmark, although performance at 1% FAR remains limited.

Together, these results support improved RS singer verification under globally singer- and singer-specific-checkpoint-disjoint evaluation, with improvements retained on external zero-day songs.

4. Manuscript updates and individual responses

These experiments are complete. We will update dataset statistics, experimental descriptions, tables and related claims throughout the manuscript for consistency. Additional analyses and reviewer-specific responses appear in our individual replies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.