{"id":"37aa365f-7b44-4831-9986-d3895f385acc","arxiv_id":"2508.08924","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"The abstract claims EGGCodec reduces F0 MAE from 14.14 Hz to 13.69 Hz, but the manuscript body omits the EGGCodec content entirely.","lead":"This preprint claims a neural codec for electroglottography signals that reconstructs the signal and extracts vocal pitch. The provided full text is an unrelated paper on light field display blur, so the EGGCodec results cannot be verified.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Submission's full text is an unrelated paper; the EGGCodec central claim has no supporting evidence in the provided manuscript.","rationale":"The reader's weakest assumption is that the abstract's premise — reconstructed EGG signals correspond more closely to F0 than conventional Encodec features — is not demonstrated. My stress-test identifies an even more fundamental and load-bearing issue: the entire manuscript body is an unrelated paper. Every part of the abstract's central claim depends on experiments, ablations, and architectural details that are absent. This is not a matter of an unproven hypothesis; it is a complete absence of supporting evidence. Therefore the honest verdict is unchanged: the claims remain unverified. I agree with the reader's assessment and with the UNVERDICTED verdict, and no adjustment is needed beyond maintaining that status. If a correct manuscript is later provided, the concrete test above would give a straightforward path toward verification.","tokens_in":14196,"tokens_out":2430,"duration_ms":26413,"concrete_test":"Download the PDF for arXiv:2508.08924 and compare its body text to the abstract. If the body is the DASC light-field paper, the EGGCodec claims are unverifiable and the submission must be returned for the correct manuscript. If a corrected EGGCodec manuscript is supplied, independently rerun the F0 extraction evaluation on the stated EGG-inclusive dataset using the released code/checkpoints and check whether MAE drops from 14.14 Hz to 13.69 Hz and VDE improves by 38.2%.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract — that EGGCodec reduces F0 extraction MAE from 14.14 Hz to 13.69 Hz and improves VDE by 38.2% — is presented as the result of a concrete architecture, a multi-scale frequency-domain loss, a time-domain correlation loss, and ablation experiments. For this claim to be verifiable, the submitted full text must describe that architecture, those losses, and those experiments. Instead, the full text is arXiv:2508.08928, 'DASC: Depth-of-Field Aware Scene Complexity Metric for 3D Visualization on Light Field Display', a paper about light field displays containing no EGG signals, no Encodec framework, no F0 extraction, and no ablation study. The abstract's secondary mechanism claim — 'reconstructed EGG signals, which more closely correspond to F0' — is likewise untestable because no supporting analysis, comparison, or derivation appears in the body. In short, the manuscript as submitted contains zero evidence for its headline result; the claim is unsupported rather than merely weakly supported. This is a submission-level consistency failure, not a technical critique of the proposed method, but it blocks any scientific evaluation of the abstract's assertions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The abstract of arXiv:2508.08924 describes EGGCodec, a neural Encodec framework for electroglottography (EGG) signal reconstruction and F0 extraction, proposing a multi-scale frequency-domain loss and a time-domain correlation loss, and reporting that EGGCodec reduces F0 MAE from 14.14 Hz to 13.69 Hz and improves VDE by 38.2% relative to state-of-the-art schemes. The submitted full text, however, is an entirely different paper titled 'DASC: Depth-of-Field Aware Scene Complexity Metric for 3D Visualization on Light Field Display' (arXiv:2508.08928), which deals with light field displays and contains no EGG signals, no Encodec architecture, no F0 extraction, and no ablation experiments for EGGCodec. Consequently, none of the abstract's claims can be checked against the body of the manuscript.","tokens_in":14394,"tokens_out":3183,"duration_ms":31925,"significance":"If the reported results were substantiated, EGGCodec would be a competently engineered contribution to speech analysis, with a modest absolute improvement in F0 MAE and a large relative VDE improvement, plus a clean design insight (using reconstructed EGG rather than codec features for F0). The paper currently provides no such substantiation: there is no architecture description, no training or evaluation protocol, no dataset specification, no statistical analysis, and no code. The significance of the claimed contribution therefore cannot be assessed from the submitted manuscript.","major_comments":[{"comment":"The full text is the DASC paper on light field displays, not the EGGCodec paper announced in the abstract; it contains no mention of electroglottography, Encodec, F0, or the proposed losses. As a result, the central quantitative claims in the abstract—MAE reduction from 14.14 Hz to 13.69 Hz and VDE improvement of 38.2%—are entirely unsupported by experimental evidence in the manuscript, and the method cannot be reproduced or evaluated.","section":"Full text (all sections)"},{"comment":"The abstract asserts that 'reconstructed EGG signals, which more closely correspond to F0' explain the performance gain, but this premise is not demonstrated anywhere in the submitted text; no comparison between reconstructed-EGG features and conventional Encodec features is provided. This leaves the causal story of the claimed improvement untestable.","section":"Abstract, mechanism claim"}],"minor_comments":[{"comment":"Because the full text is a different paper, its equations, references, and subjective-study descriptions are irrelevant to the abstract's claims; a reader cannot locate any of the EGGCodec components named in the abstract.","section":"Full text, general"},{"comment":"These equations pertain to light field display geometry and Bradley-Terry preference analysis, not to neural audio coding or F0 extraction, so they cannot be used to verify the claimed ablation results.","section":"Full text, Eq. (1)–(26)"}],"recommendation":"reject","confidential_remarks":"The disparity between the abstract and the full text suggests an administrative submission mismatch, but as submitted the manuscript is not reviewable for the claims it announces. If the authors intended to submit the EGGCodec paper, they should withdraw this version and resubmit the matching manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the submission is broken. The abstract advertises EGGCodec, a neural codec for EGG signal reconstruction and F0 extraction, but the full text is an unrelated paper about depth-of-field aware scene complexity for light field displays (arXiv:2508.08928), with a different author list. As submitted, this manuscript contains zero evidence for its headline claims. There is no way to review it as is.\n\nWhat's actually new: the abstract describes a plausible extension of the Encodec architecture to EGG signals, with a multi-scale frequency-domain loss and a time-domain correlation loss, and it proposes dropping the GAN discriminator. If that design works as claimed, it could be a useful incremental contribution for the speech/voice community. But the claimed results—MAE reduction from 14.14 Hz to 13.69 Hz, VDE improvement of 38.2%—are only numbers in the abstract. The body text contains no experiments, no dataset, no ablation study, no architecture details, and no statistical analysis. There is nothing to check.\n\nThe full text that is attached does look like a normal paper (a subjective study on light field displays with a proposed DASC metric and a validation study). That may be a decent submission to IEEE TMM, but it is not this paper. The author list mismatch (Akbar et al. versus Feng et al.) makes it plain that the wrong file was submitted rather than a disguised merge.\n\nThe soft spot is not small; it is load-bearing. The central scientific claim has no supporting evidence in the manuscript. I cannot evaluate the method's soundness, novelty, or the credibility of the reported numbers. Even if the right paper were attached, the MAE gain is slight and the significance would probably be incremental, but at least a referee could judge the architecture and the loss design. As it stands, the submission is an abstract with a photocopied article behind it.\n\nWho benefits: nobody, as submitted. If the authors resubmit the correct full text, this becomes a reviewable paper that deserves a serious referee for the EGG/F0 community. The idea of using reconstructed EGG signals for F0 extraction rather than features from a codec latent is worth testing, and removing GAN training is a practical simplification. But that evaluation has to wait for the actual content.\n\nRecommendation: desk reject this version. Tell the authors to resubmit with the correct manuscript. Do not send the current file to peer review.","headline":"Submission is a file mix-up: the abstract describes EGGCodec, the full text is an unrelated light field display paper, and none of the claimed F0 results have any supporting evidence.","tokens_in":14854,"tokens_out":3248,"would_cite":false,"duration_ms":32156,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural codec that reconstructs EGG signals extracts F0 with lower error than existing F0 schemes.","keywords":["electroglottography","EGG signal reconstruction","fundamental frequency extraction","F0 estimation","Encodec","multi-scale frequency-domain loss","time-domain correlation loss","voicing decision error"],"falsifier":"Retrain EGGCodec twice: once extracting F0 from the reconstructed EGG signal and once extracting F0 from the encoder's feature vector, keeping all other components identical. If the feature-based variant matches or beats the reconstruction-based variant on mean absolute F0 error and voicing decision error, the paper's central claim is false.","tokens_in":14042,"feed_emoji":"🎤","tokens_out":5080,"duration_ms":46913,"temperature":0.7,"pith_summary":"This letter tries to show that a neural codec built for electroglottography (EGG) signals can both reconstruct the EGG waveform and extract the fundamental frequency (F0) more reliably than current F0 extraction systems. The proposed framework, EGGCodec, adds a multi-scale frequency-domain loss and a time-domain correlation loss to an Encodec-style architecture, and deliberately removes the usual GAN discriminator. Evaluated on a widely used EGG-inclusive dataset, it reports a fall in mean absolute F0 error from 14.14 Hz to 13.69 Hz and a 38.2% improvement in voicing decision error. The practical appeal is that one model could serve both speech coding and F0-related downstream tasks, with F0 taken from the reconstructed EGG signal rather than from abstract codec features.","feed_headline":"Neural codec cuts F0 error by rebuilding EGG signals","feed_subtitle":"EGGCodec reports F0 mean absolute error at 13.69 Hz and a 38.2% voicing-decision-error improvement.","key_machinery":"The load-bearing mechanism is a pair of training losses attached to an Encodec-style encoder-decoder: a multi-scale frequency-domain loss that compares original and reconstructed EGG spectra at several scales, and a time-domain correlation loss that aligns the waveforms' correlation structure. The architectural twist is that F0 is estimated from the decoder's reconstructed EGG signal, not from the compressed feature space, and the standard GAN discriminator is removed. Together these components make the reconstruction faithful enough to serve as the substrate for F0 extraction, which the paper claims is the reason for the accuracy gain.","core_discovery":"The central claim is that F0 extraction is more accurate when the task is reframed as EGG reconstruction: instead of reading F0 off the encoder features of a conventional Encodec, EGGCodec first reconstructs the EGG signal and then performs F0 extraction on that reconstruction. The paper argues that reconstructed EGG signals 'more closely correspond to F0' than the features used by conventional models, and attributes the gains to a multi-scale frequency-domain loss that captures the relationship between original and reconstructed EGG, plus a time-domain correlation loss that improves generalization. It also claims that dropping the GAN discriminator simplifies training with only negligible performance loss. On its evaluation set, these choices yield a mean absolute error reduction from 14.14 Hz to 13.69 Hz and a 38.2% relative improvement in voicing decision error against state-of-the-art F0 extraction schemes.","pith_inferences":["If the reconstruction-first principle is the true source of the gain, then any codec or vocoder that outputs a waveform matched to the physical glottal signal could be used as a front end for F0 extraction, not just Encodec-style architectures.","A natural extension is to test whether the same multi-scale frequency-domain plus correlation-loss recipe improves F0 from reconstructed speech waveforms rather than EGG, which would widen its applicability to ordinary audio.","One could also probe the causal link directly by monitoring F0 error as reconstruction fidelity is artificially degraded; if F0 error tracks reconstruction error, the paper's assertion would be supported."],"forward_implications":["If EGGCodec works as reported, a single Encodec-style model can compress EGG signals and deliver F0 estimates at the same time, simplifying pipelines that today separate coding and F0 analysis.","The 38.2% relative gain in voicing decision error suggests the reconstructed EGG preserves voicing onset and offset information better than feature-based F0 estimators, which matters for prosody and expressive speech processing.","Removing the GAN discriminator makes training lighter, so the approach could transfer more easily to smaller, domain-specific EGG datasets than adversarial codecs.","The reported MAE drop from 14.14 Hz to 13.69 Hz, while small in absolute terms, would be most consequential in low-F0 or high-precision settings where a few Hertz separate acceptable from poor estimates."],"supporting_citations":[],"fun_headline_variants":["EGGCodec: Rebuild EGG, refine F0","F0 extraction via EGG reconstruction, not features","Drop the GAN, boost F0 with EGG reconstruction","EGG reconstruction improves F0 extraction to 13.69 Hz","Reframe F0 as EGG reconstruction for accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a reconstructed EGG waveform carries F0 information more faithfully than the latent features of a conventional Encodec; the available text asserts this rather than demonstrates it, and the reported improvement would not follow if the reverse were true.","fun_headline_variants_meta":{"raw":{"variants":["EGGCodec: Rebuild EGG, refine F0","F0 extraction via EGG reconstruction, not features","Drop the GAN, boost F0 with EGG reconstruction","EGG reconstruction improves F0 extraction to 13.69 Hz","Reframe F0 as EGG reconstruction for accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000438,"raw_usage":{"total_tokens":2215,"prompt_tokens":924,"completion_tokens":1291,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":1207}},"tokens_in":540,"tokens_out":1291,"duration_ms":12087,"temperature":1.0,"reasoning_tokens":1207,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:31:28.789634+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain EGGCodec twice: once extracting F0 from the reconstructed EGG signal and once extracting F0 from the encoder's feature vector, keeping all other components identical. If the feature-based variant matches or beats the reconstruction-based variant on mean absolute F0 error and voicing decision error, the paper's central claim is false.","supporting_citations":[],"review_version":1}