{"id":"77b91da1-b8a6-4d35-bb08-ad61aec4ecec","arxiv_id":"2412.17796","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"x-vector embeddings and a Rényi divergence fusion loss achieve the best audio deepfake source attribution on ASVspoof 2019 and CFAD, though the benchmark protocol is non-standard.","lead":"This paper tests six speech pre-trained models for audio deepfake source attribution, identifying which AI voice generator made a clip, and proposes a fusion method called FINDER that combines two models with a Rényi divergence loss. The best combination reaches high accuracy on two benchmarks, but the evaluation protocol is non-standard and error bars are missing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ASV SOTA and x-vector prosody claims rest on a non-speaker-disjoint 5-fold split, so speaker/recording leakage may inflate accuracy and confound the prosody explanation.","rationale":"The reader's weakest_assumption correctly pinpoints the non-speaker-disjoint random split of ASV as the most load-bearing issue, and I agree with that identification. The paper's central claims are empirical: x-vector is the best individual PTM because speaker-recognition pretraining captures prosodic signatures, and FINDER fusion of Whisper+x-vector attains SOTA. Both claims depend on the integrity of the evaluation protocol. Without speaker-/recording-disjoint folds, the reported ASV numbers can be inflated by identity leakage, and the mechanism attributed to prosody is confounded by speaker identity. The paper's own admission that direct comparison to prior work is impossible further weakens the SOTA claim. I do not see this as a reason to reject the paper outright: the fusion finding is internally compared against concatenation under the same protocol, and the CFAD results use the official split, so at least part of the empirical contribution may survive. The correct response is to require the speaker-disjoint rerun and a direct evaluation on the official ASV eval split before accepting the headline numbers or the prosody interpretation. The reader already issued a CONDITIONAL verdict; my stress test reinforces that conditional, rather than moving it to another verdict. The RD-loss concern noted in the rationale is secondary because the empirical comparison of FINDER vs. concatenation is direct, whereas protocol leakage contaminates the central claim's interpretation and comparability. Overall, I found no additional load-bearing flaw beyond the one the reader identified; the fix is concrete and experimentally feasible.","tokens_in":10112,"tokens_out":4989,"duration_ms":52855,"concrete_test":"Rerun all ASV experiments in Tables 1 and 2 using group 5-fold cross-validation where groups are ASVspoof speaker IDs (and, more strictly, original utterance IDs), ensuring no speaker or source recording appears in both training and test folds; report accuracy and EER per fold. As a second check, train the Whisper+x-vector FINDER model on ASV train+dev only and evaluate on the official ASV eval partition to enable direct comparison with prior SOTA. If x-vector's margin over Whisper drops substantially or the FINDER accuracy falls by more than about 5 points under the speaker-disjoint protocol, the x-vector-prosody and SOTA conclusions are inflated by leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 says: 'We combine the train, validation and testing splits for ASV' and 'followed 5-fold cross validation,' with no statement that folds are speaker-disjoint or recording-disjoint. ASVspoof 2019 LA was built with distinct speakers across its original train/dev/eval partitions; pooling those partitions and randomly splitting utterances almost guarantees that the same speaker appears in both training and test folds. Moreover, spoofed utterances across different A-classes are generated from shared bonafide source recordings, so the same utterance content can appear in both train and test folds, even under different source classes. The downstream CNN, especially on x-vector features (a speaker embedding), can then exploit speaker or utterance identity as a shortcut. This directly undermines the central claim in two ways. First, the reported 98.91% accuracy / 0.26% EER on ASV is not an honest source-attribution number; the paper itself concedes 'we can't directly compare our results to previous works' yet still claims SOTA in the abstract, conclusion, and Table 3. Second, the explanation that x-vector outperforms other PTMs because speaker-recognition pre-training captures prosodic source signatures is confounded: under leakage, x-vector's speaker-identity features are the most direct cue, independent of prosody. The fusion result of Whisper+x-vector may survive a cleaner protocol, but the absolute numbers and the prosodic interpretation are not trustworthy as reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates speech pre-trained models (PTMs) for audio deepfake source attribution (ADSD), focusing on prosodic signatures of generative sources. It compares six PTMs (wav2vec2, wav2vec2-emo, XLS-R, Whisper, WavLM, x-vector) with FCN and CNN downstream networks on ASVspoof 2019 and CFAD. It further proposes FINDER, a fusion method that adds a Rényi-divergence loss between projected PTM representations to the cross-entropy loss. The main empirical claims are: (i) x-vector is the best individual PTM, attributed to speaker-recognition pre-training capturing prosodic characteristics; (ii) fusion of Whisper and x-vector representations through FINDER achieves the topmost performance, with 98.91% accuracy / 0.26% EER on ASV and 99.01% accuracy / 1.07% EER on CFAD, described as state-of-the-art (SOTA).","tokens_in":10537,"tokens_out":3444,"duration_ms":33245,"significance":"The paper addresses a relevant and timely problem, audio deepfake source attribution, and provides a broad comparison of established PTMs. Its proposed fusion loss is simple and the experimental matrix is internally consistent. The strongest contribution is the empirical evidence that fusion of complementary PTM representations improves attribution over individual PTMs, and that FINDER consistently outperforms concatenation baselines under the same evaluation setup. The paper also ships useful t-SNE visualizations and promises code release. However, the central quantitative claims and the prosodic interpretation rest on an evaluation protocol that is not speaker-disjoint or recording-disjoint, which calls into question the absolute accuracy/EER numbers and the SOTA label. If the protocol issue is fixed and the results remain qualitatively similar, the paper would be a worthwhile contribution; in its current form, the main claims are not trustworthy as stated.","major_comments":[{"comment":"The ASV evaluation protocol is the load-bearing element of the paper's SOTA claim. The text states: 'We combine the train, validation and testing splits for ASV' and 'We followed 5-fold cross validation for ASV,' with no mention of speaker-disjoint or recording-disjoint folds. ASVspoof 2019 LA was constructed with distinct speakers across its original partitions, and the spoofed A-class samples are generated from shared bonafide source utterances. Randomly splitting the pooled utterances into 5 folds therefore makes it very likely that the same speaker, and even the same utterance content, appears in both training and test folds. The downstream CNN, particularly on x-vector speaker embeddings, can exploit this identity leakage as a shortcut. This inflates the reported 98.91% accuracy / 0.26% EER and confounds the attribution interpretation. The authors must either enforce speaker/recording-disjoint folds (e.g., by grouping utterances by source recording and speaker) and report results under that protocol, or explicitly justify why random folds constitute a valid attribution evaluation.","section":"Section 5, Benchmark Datasets; Section 5, Comparison to Previous Works; Table 3"},{"comment":"The paper states 'as we have considered all the source classes across train, validation, and test split for ASV, so we can't directly compare our results to previous works,' yet the abstract, conclusion, and Table 3 still claim SOTA performance. The reimplemented baselines (MiO and AASIST) in Table 3 are evaluated under the same non-standard protocol, so the relative comparison is internally consistent; however, the claim of 'attaining SOTA performance' against the published literature is unsupported. Please either temper the claim to 'SOTA under our evaluation protocol' or provide a protocol-matched comparison using published numbers where the original train/dev/test split is respected.","section":"Section 5, Comparison to Previous Works; Table 3; Abstract; Conclusion"},{"comment":"The interpretation that x-vector outperforms other PTMs because its speaker-recognition pre-training captures prosodic source signatures is confounded by the protocol leakage. Under a random split of pooled utterances, x-vector embeddings give the most direct access to speaker identity, which is an even simpler cue than prosody. To support the prosody-based explanation, the authors need to control for speaker/recording overlap, for example by evaluating on speaker-disjoint folds or by ablating speaker-related dimensions of the x-vector embedding. Without such a control, the observed x-vector advantage cannot be attributed to prosodic signature capture.","section":"Section 5, Experimental Results; Table 1; Section 6, Conclusion"},{"comment":"The hyperparameters α, ε, and λ were chosen 'through some preliminary exploration' and the best fusion pair (Whisper + x-vector) is the selected maximum over the fourteen fusion pairs in Table 2. The reported 98.91% and 99.01% numbers are therefore a selected maximum over hyperparameters and fusion pairs, not an independent out-of-sample prediction. The authors should describe the model selection procedure (e.g., a separate validation split) or at least report the variance across the 5 folds and disclose that the best pair is chosen on the evaluation set.","section":"Section 5, Training Details; Table 2"}],"minor_comments":[{"comment":"The Rényi divergence loss adds a small constant ε inside both (ea,i+ε) and (eb,i+ε) before the log and power operations; please clarify whether this shift is intended to bias the divergence estimate or solely to ensure numerical stability, and state how ε interacts with the choice of α.","section":"Section 4, FINDER equation"},{"comment":"The caption of Table 2 says 'All the scores are average of 5-folds,' but the text says CFAD uses the 'official split given for CFAD.' Please clarify whether CFAD results are also averaged over folds or are single-run numbers from the official split.","section":"Section 5, Benchmark Datasets; Table 2 caption"},{"comment":"The sentence 'We extract representations from the last hidden state of the PTMs by average pooling. We' is truncated; it should be completed or removed.","section":"Appendix A.1, last paragraph"},{"comment":"There is an unmatched opening parenthesis in 'max pooling (pool size 2.' and the following sentence also has a typo ('succeeded by followed by'). Please fix these punctuation errors.","section":"Appendix A.3, System Configurations"},{"comment":"The phrase 'theFINDER' appears without a space at the end of the comparison paragraph; also the capitalization 'FINDER (FusIon through Re Nyi DivERgence)' is unconventional and should be normalized.","section":"Section 5, Comparison to Previous Works"},{"comment":"The EER computation is described only as 'the average scores of one-vs-all.' Please specify how the one-vs-all EER is aggregated (e.g., macro-average over classes) and whether the EER is computed on scores or on probabilities, since this affects comparability with other work.","section":"Section 5, Experimental Results"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is reasonable and the fusion method is simple to reproduce. The main issue is the evaluation protocol: the ASV results and the SOTA claim rest on a non-speaker-disjoint random 5-fold split, and the paper itself concedes that direct comparison with prior work is impossible. This is fixable within the scope of the manuscript by enforcing speaker/recording-disjoint folds and re-phrasing the SOTA claim. The prosodic interpretation also needs a control for speaker identity. I recommend major revision rather than rejection, because the relative comparisons within the paper (FINDER vs. concatenation, and the matrix of PTM pairs) are informative and likely to survive a cleaner protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a useful empirical sweep with a load-bearing evaluation flaw. The authors compare six pre-trained speech models for audio deepfake source attribution and propose a fusion loss, FINDER. That comparison is genuinely new for this task, and FINDER beats plain concatenation in most pairs they tested. Credit where it's due: they ran six PTMs on two datasets, reported accuracy and EER, and included t-SNE visualizations. The x-vector result is interesting, but the prosody explanation is not supported by the evidence as presented.\n\nThe soft spot is in Section 5. They pool ASVspoof 2019 train/val/test splits and do random 5-fold CV with no speaker or recording disjointness. ASVspoof 2019 was built with distinct speakers across its original partitions; pooling and randomly splitting almost guarantees the same speaker appears in both train and test folds. On top of that, spoofed utterances across different classes are generated from shared bona fide source recordings, so utterance-level leakage is plausible. Since x-vector is a speaker embedding, the downstream CNN can exploit speaker identity as a shortcut. That inflates the headline 98.91% accuracy / 0.26% EER and makes the prosody interpretation untestable. The paper itself says \"we can't directly compare our results to previous works,\" yet the abstract and conclusion claim SOTA. That is inconsistent.\n\nThe stress-test note is right. Also, the best fusion pair (Whisper + x-vector) is selected from fourteen pairs, and hyperparameters (α, λ, ε, D) were tuned on the same evaluation. No error bars, no code release yet. Rényi divergence is applied to raw feature vectors, not probability distributions, so the theoretical grounding is loose. These are fixable: speaker-disjoint folds, confidence intervals, and a clear statement that the reported best numbers are selected maxima. The CFAD results use the official split and are less affected, but the same selection-bias caveat applies.\n\nFor a reader working on audio deepfake attribution, this paper is worth reading as a benchmark comparison and as a cautionary example of evaluation pitfalls. It deserves a serious referee, but the revision should require speaker-disjoint evaluation, honest SOTA claims, and preferably code release before the numbers are taken at face value.","headline":"A useful PTM comparison for audio deepfake source attribution, but the ASVspoof protocol leaks speaker identity and the SOTA claim does not survive a speaker-disjoint evaluation.","tokens_in":10939,"tokens_out":1676,"would_cite":false,"duration_ms":17328,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fusing Whisper and x-vector embeddings with a Rényi-divergence loss attributes audio deepfakes to their generating system, reaching 98.91% accuracy on ASVspoof 2019 and 99.01% accuracy on CFAD.","keywords":["audio deepfake detection","source attribution","prosodic signatures","pre-trained speech models","x-vector","Rényi divergence","feature fusion","ASVspoof 2019"],"falsifier":"Re-run the exact FINDER (Whisper + x-vector) pipeline on ASVspoof 2019 and CFAD with speaker-disjoint and recording-disjoint folds in place of the random split, and compare accuracy and equal error rate to 98.91%/0.26% and 99.01%/1.07%; a large rise in EER would show the headline result depends on the optimistic split rather than on source prosody.","tokens_in":9913,"feed_emoji":"🎙️","tokens_out":11199,"duration_ms":96287,"temperature":0.7,"pith_summary":"Audio deepfakes carry prosodic fingerprints—pitch, tone, rhythm, intonation—that differ from one text-to-speech or voice-conversion generator to the next. The paper asks which speech pretrained models best expose those fingerprints for source attribution, and finds that x-vector, a compact speaker-recognition embedding, outperforms larger general-purpose models such as Whisper and XLS-R. It then proposes FINDER, a fusion method that uses Rényi divergence to align two pretrained representation spaces while training a classifier, and reports that Whisper plus x-vector fused this way gives the best closed-set attribution: 98.91% accuracy with 0.26% equal error rate on ASVspoof 2019 and 99.01% accuracy with 1.07% equal error rate on CFAD. If the result holds, forensic attribution of a deepfake to its generating tool can be done by freezing pretrained speech embeddings and training only a small head.","feed_headline":"Fusing two speech embeddings attributes deepfake audio to its source","feed_subtitle":"Whisper + x-vector embeddings aligned by Rényi divergence reach 98.91% and 99.01% accuracy on the two benchmarks.","key_machinery":"Two mechanisms carry the argument. First, x-vector is a time-delay neural network embedding trained for speaker recognition; the paper treats its learned speaker-prosody sensitivity as a transferable fingerprint extractor for generator identity. Second, FINDER fuses two frozen pretrained-model representation spaces by projecting each through convolutional blocks, then adding a Rényi-divergence loss $L_{\\mathrm{RD}} = \\frac{1}{\\alpha-1}\\log \\sum_i (e_{a,i}+\\epsilon)^\\alpha(e_{b,i}+\\epsilon)^{1-\\alpha}$ to the cross-entropy loss. The divergence term pulls the two representations toward each other while the classifier separates generator classes, with hyperparameters $\\alpha=2$, $\\epsilon=0.1$, and $\\lambda=0.4$ weighting the two losses. This alignment is what lets the complementary ASR-style information of Whisper and speaker-style information of x-vector combine more effectively than concatenation.","core_discovery":"On the paper's own terms, the central discovery is that prosodic source signatures are best captured by a speaker-recognition pretrained model, x-vector, and that fusing those embeddings with Whisper's via the proposed FINDER loss is the strongest configuration. In controlled comparisons across six pretrained models and two downstream heads, x-vector needed only about 4.2 million parameters yet beat larger monolingual and multilingual models on both benchmark datasets, which the authors attribute to speaker-recognition pretraining. Replacing simple concatenation with their Rényi-divergence fusion consistently improved results for nearly every pretrained-model pair. The best pair, Whisper and x-vector, reaches 98.91% accuracy with 0.26% equal error rate on ASVspoof 2019 and 99.01% accuracy with 1.07% equal error rate on CFAD, ahead of the reimplemented MiO and AASIST baselines. Those numbers are closed-set: the system can only attribute audio to generator classes seen in training.","pith_inferences":["The paper does not enforce speaker-disjoint or recording-disjoint folds on ASVspoof 2019, so the reported 98.91% accuracy and 0.26% EER could be lower under a split that prevents the same speaker or recording from appearing in both training and test folds.","A testable extension, not in the paper, is adding explicit prosody features such as pitch, energy, or rhythm to the FINDER alignment; this would show whether prosody itself, rather than channel or recording artifacts, drives the result.","The Rényi-divergence fusion loss is model-agnostic and could be evaluated as a generic multi-encoder fusion loss for other speech tasks such as emotion recognition or speaker verification.","The paper's own limitations section notes the closed-vocabulary setting; holding out entire generator classes would test whether the approach can attribute audio to TTS or VC systems never seen in training."],"forward_implications":["Closed-set attribution of deepfake audio can be built from frozen pretrained embeddings plus a small trainable classifier, without training a large audio model from scratch.","Speaker-recognition pretraining appears more useful for prosodic source fingerprints than much larger self-supervised speech models; model size is not the controlling factor.","Fusing an ASR-oriented model with a speaker-oriented model through divergence alignment beats concatenation of the same embeddings.","The paper gives the first audio-deepfake source-attribution numbers on CFAD and a new baseline for ASVspoof 2019 source attribution.","The system does not generalize to generators absent from the training set; open-vocabulary attribution remains unsolved."],"supporting_citations":[{"why":"Supplies the x-vector speaker-recognition embedding that is the best individual representation in the paper.","marker":"Snyder et al., 2018"},{"why":"Supplies Whisper, the ASR pretrained model whose fusion with x-vector gives the top attribution results.","marker":"Radford et al., 2023"},{"why":"Gives the Rényi-divergence definition that FINDER turns into a feature-alignment loss.","marker":"Van Erven and Harremos, 2014"},{"why":"Provides the ASVspoof 2019 dataset with its 19 spoofed source classes used for attribution.","marker":"Wang et al., 2020"},{"why":"Provides CFAD, the Chinese fake-audio dataset for which the paper reports first source-attribution results.","marker":"Ma et al., 2024"},{"why":"Prior pretrained-model-based source-tracing work the paper extends, and whose AASIST-with-Wav2vec2 setup is reimplemented as a baseline.","marker":"Klein et al., 2024a"},{"why":"Motivates pretrained-model representation fusion and supplies the MiO baseline reimplemented with Whisper and x-vector.","marker":"Chetia Phukan et al., 2024"},{"why":"Supplies AASIST, the downstream architecture used as a comparison baseline on both datasets.","marker":"Jung et al., 2022"},{"why":"Defines Wav2vec2, one of the monolingual pretrained models included in the comparative study.","marker":"Baevski et al., 2020"}],"fun_headline_variants":["Speaker model keys prosody for deepfake source attribution","Whisper + x-vector fusion attributes deepfake audio best","Prosody via speaker pretraining pinpoints deepfake origin","Fusing Whisper and x-vector tops deepfake source attribution","x-vector prosody beats bigger models for deepfake source"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the random five-fold split of the combined ASVspoof 2019 partitions, without speaker-disjoint or recording-disjoint folds, is a valid measure of source attribution; if that split is optimistic, the reported numbers are inflated.","fun_headline_variants_meta":{"raw":{"variants":["Speaker model keys prosody for deepfake source attribution","Whisper + x-vector fusion attributes deepfake audio best","Prosody via speaker pretraining pinpoints deepfake origin","Fusing Whisper and x-vector tops deepfake source attribution","x-vector prosody beats bigger models for deepfake source"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00068,"raw_usage":{"total_tokens":3119,"prompt_tokens":1001,"completion_tokens":2118,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":2043}},"tokens_in":617,"tokens_out":2118,"duration_ms":15603,"temperature":1.0,"reasoning_tokens":2043,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:08:19.938366+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the exact FINDER (Whisper + x-vector) pipeline on ASVspoof 2019 and CFAD with speaker-disjoint and recording-disjoint folds in place of the random split, and compare accuracy and equal error rate to 98.91%/0.26% and 99.01%/1.07%; a large rise in EER would show the headline result depends on the optimistic split rather than on source prosody.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ASVspoof 2019 dataset with its 19 spoofed source classes used for attribution."}],"review_version":1}