{"id":"62f4e117-7511-454e-81a7-11235d012e04","arxiv_id":"2509.09489","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An audio-only speech inversion system, fine-tuned on children with velopharyngeal insufficiency, estimates nasalance with improved correlation over a prior adult baseline.","lead":"Researchers built a neural network that estimates nasalance, the amount of nasal airflow during speech, from a plain audio recording, first training on adults and then fine-tuning on children with velopharyngeal insufficiency. If the method holds up in larger clinical data, it could give doctors a non-invasive way to monitor this speech disorder in children.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline PPMC gains are not shown to exceed speaker-level noise: adult comparison rests on 2 test speakers; VPI fine-tuning gain is smaller than reported fold-to-fold SD.","rationale":"The reader's weakest assumption—that the small test sets and lack of significance testing leave the headline PPMC improvements vulnerable to speaker-specific noise—is exactly the most load-bearing concern. The paper's value depends on those numeric gains being real, but the evidence provided is insufficient to rule out chance. I considered other potential issues: the normalization across all utterances within a dataset might seem like leakage, but PPMC is invariant to positive affine transformations, so min-max scaling that uses test-set statistics does not by itself inflate the correlation metric; thus it is not a decisive concern. The conflation of architecture, SSL model, and segment-length changes in the adult comparison is real but secondary, because the paper also reports an ablation study and a comparison against a retrained baseline on the same data. The lack of released code/data is a reproducibility issue, not a direct challenge to the internal validity of the reported numbers. The most direct way to destabilize the central claim is to show that the reported PPMC differences are within noise; the proposed bootstrap and per-speaker/per-fold analysis would settle exactly that. Therefore I agree with the reader's conditional verdict and would not change it without the additional statistical evidence.","tokens_in":9685,"tokens_out":5686,"duration_ms":68495,"concrete_test":"Compute bootstrap confidence intervals for both key PPMC differences: (a) on the adult test set, resample the 70 utterances or the 2 speakers with replacement (10,000 iterations) to estimate the 95% CI for the WavLM multi-task vs HuBERT baseline VP PPMC difference, and also report the VP PPMC for each test speaker separately; (b) on the VPI-child 5-fold CV, report the five individual fold VP PPMC values for SI and SI-FT and perform a paired test (e.g., Wilcoxon signed-rank or paired bootstrap) on the fold-level differences. If the adult CI includes zero, the improvement is not directionally consistent across both speakers, or the VPI fold-level difference is not significant at p < 0.05, then the central claims of 16.92% and 7.90% relative improvement are not statistically established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's two central quantitative claims are the 16.92% relative PPMC improvement over the [5] baseline on the healthy-adult test set and the 7.90% relative improvement from fine-tuning on VPI children. Both rest on extremely small test sets: the adult test set is only 2 speakers / 70 utterances (Table I), and each VPI-child test fold has only 3 subjects. No confidence intervals or significance tests are reported. For the VPI fine-tuning result, Table IV reports the before/after VP PPMC as 0.6357 (SD 0.14) and 0.6859 (SD 0.08); the mean difference is 0.0502, which is within the reported standard deviation of the pre-fine-tuned model and only about 0.6 of the post-fine-tuned SD. With only 5 folds and 3 test speakers per fold, this difference could easily arise from speaker-specific or fold-specific noise. Similarly, the adult improvement from 0.8115 to 0.9488 could be driven by one of the two test speakers; no per-speaker scores are given. Without a measure of uncertainty, the claimed improvements cannot be distinguished from chance. This is load-bearing because the abstract and conclusions explicitly claim 'significant' improvement and the clinical utility of the SI-FT system depends on a real gain over the non-fine-tuned model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a speech-inversion (SI) system that estimates nasalance (treated as a VP tract variable, VP TV) from audio, using self-supervised learning (SSL) embeddings (HuBERT-Large, wav2vec 2.0-Large, WavLM-Large), a multi-task objective that jointly estimates VP TV, an EGG-envelope proxy, and three glottal source features (F0, periodicity, aperiodicity), and variable-length training segments. The system is first trained on healthy-adult speech and then fine-tuned on a small dataset of children with velopharyngeal insufficiency (VPI). The paper reports a 16.92% relative PPMC improvement over the prior SI system of Siriwardena et al. [5] on the healthy-adult test set, and a 7.90% relative PPMC improvement after fine-tuning on VPI-child data, with additional qualitative cross-corpus and nasal-emission analyses.","tokens_in":10055,"tokens_out":6141,"duration_ms":77904,"significance":"If the reported gains are reliable, the work has clear clinical value: it is, to my knowledge, the first SI system aimed at estimating nasalance directly in children with VPI, and it provides a non-invasive alternative to nasometry and nasopharyngoscopy. The use of WavLM-Large embeddings for nasalance estimation, the multi-task auxiliary glottal-feature objective, and the adult-to-child fine-tuning strategy are all reasonable extensions of prior work. The paper has several good practices: it evaluates on the same healthy-adult test set as [5], it reports a retrained baseline on the current training data, and it includes an ablation of the auxiliary tasks. However, the central quantitative claims currently rest on very small test sets with no uncertainty quantification, and the headline adult improvement mixes multiple changes at once. These issues are fixable with additional analyses, but they are load-bearing for the paper's main conclusions.","major_comments":[{"comment":"The headline 16.92% relative improvement is not a controlled comparison. It compares the proposed WavLM system to the baseline result originally reported in [5], which was trained on the older, smaller healthy-adult corpus, whereas the proposed system is trained on the enlarged corpus (four additional subjects). The more controlled comparison in the same table—the same baseline architecture retrained on the current training data—yields only a 6.56% relative improvement. The abstract and conclusions nevertheless emphasize the 16.92% figure. Please make the controlled comparison the primary reported result and clearly separate the effect of added training data from the effect of the proposed model changes.","section":"§IV-A, Table II"},{"comment":"Neither of the two central claims is supported by uncertainty quantification. The healthy-adult test set consists of only 2 speakers and 70 utterances (Table I); no per-speaker or bootstrap results are given, so the gain from 0.8115 to 0.9488 could be driven by one speaker. For the VPI-child fine-tuning, Table IV reports VP PPMC increasing from 0.6357 (SD 0.14) to 0.6859 (SD 0.08) over five folds, a mean difference of 0.0502 that is smaller than the reported standard deviation of the pre-fine-tuned model. With only 3 test speakers per fold, the 7.90% relative improvement is within fold-to-fold and speaker-level noise. Please report per-fold and per-speaker scores, confidence intervals (e.g., bootstrap over speakers), and a paired significance test before claiming in the conclusions that fine-tuning 'significantly improved' nasalance estimation.","section":"§IV-A and §IV-D, Tables II and IV"},{"comment":"The claimed adult improvement conflates several simultaneous changes relative to [5]: the SSL embedding (HuBERT-Large vs. WavLM-Large), the multi-task auxiliary objective, the variable-length segment strategy, and the enlarged training corpus. Table III ablates only the auxiliary tasks within the WavLM system; it does not isolate the SSL model choice or the segment-length strategy. The conclusion that the 'proposed multi-task SI architecture and training strategy' is responsible for the improvement is therefore not supported by the experiments as presented. Please add single-factor ablations (e.g., WavLM with the baseline fixed-segment/task setup; HuBERT with the proposed multi-task variable-length setup). The ablation also shows that the glottal source features themselves contribute only a small absolute gain (VP PPMC 0.9444 without them vs. 0.9488 with them, Table III), which should be","section":"§III-B and §IV-A, Tables II and III"}],"minor_comments":[{"comment":"Typo: 'an 7.90%' should be 'a 7.90%'.","section":"Abstract"},{"comment":"The sentence 'In equation 1, α was set to 0.2' should refer to equation 2, where α is defined.","section":"§III-A, Eq. (2)"},{"comment":"Typo: 'later used to adopt the SI systems' should be 'later used to adapt the SI systems'.","section":"§III-B1"},{"comment":"State clearly whether the min-max normalization to [-1,1] is computed before or after the speaker-independent split. If computed over all utterances including the test set, this is a mild form of test-set leakage; it does not affect PPMC (which is affine-invariant) but it does affect RMSE-based training and should be described.","section":"§III-A"},{"comment":"The cross-corpus generalization claim is based on a single utterance and is only qualitative. It would be more appropriate to label this as an illustrative case study rather than a generalization evaluation, or to add a small quantitative summary over more utterances.","section":"§IV-C, Figure 3"},{"comment":"Reference [22] appears incomplete/malformed ('vol. 0, no, 1994'). Please fix.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is closely related to prior work by the same group ([5], [7], [11]) and the incremental contribution is substantial enough for a journal if the statistical concerns are addressed. The main risk is that the reported gains are within noise on the very small test sets; I would not accept the current results at face value. I would encourage the editor to request per-speaker results and significance testing, and to ask the authors to de-confound the adult comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is the first attempt I know of to estimate nasalance directly from audio in children with VPI via a speech-inversion system, and it is a sensible engineering step. The authors build on strong prior work—WavLM-Large embeddings, multi-task learning with source features, and the same test set used by Siriwardena et al. [5]—and they include a useful ablation showing that glottal auxiliary tasks help VP TV estimation. The qualitative figures, especially the nasal rustle and hypernasality examples, show the model is tracking something meaningful. The speaker-independent splits and 5-fold CV for the child data are the right instincts.\n\nThe load-bearing problem is that the two headline quantitative claims are not actually supported by the numbers as reported. The 16.92% relative improvement over [5] is on a 2-speaker adult test set and conflates the new architecture with a larger training set (the original [5] result was trained on fewer speakers); the fair comparison against a retrained baseline gives 6.56%, still on 2 speakers. Even that could be driven by one speaker—no per-speaker scores are given. For the VPI fine-tuning result, the mean PPMC gain of 0.0502 is smaller than the reported fold-to-fold SD of the pre-fine-tuned model (0.14). With 5 folds and only 3 test children per fold, that difference is well within sampling noise. No significance test or confidence interval is reported anywhere, so the abstract's word \"significant\" lacks statistical backing.\n\nThere are also minor grievances: they normalize features using whole-dataset statistics including the test split, which is sloppy though it likely doesn't affect PPMC (affine transforms leave Pearson r invariant). No code or data is released, which makes it hard to reproduce or build on. None of this is fatal to the core idea—the paper is an application first, and the from-scratch child-only comparison (0.6388 vs 0.6859) at least offers an internal sanity check. But the clinical promise depends on a real, reliable fine-tuning gain, and this evaluation does not demonstrate that yet.\n\nWho gets value from this? People working on articulatory inversion or clinical voice assessment will want to know this exists, and the method is a reasonable baseline for future work. I would encourage a serious referee to engage with it, primarily to push for per-speaker results, fold-level paired tests, and ideally an external validation set with more speakers. The field benefits from this being reviewed rather than desk-rejected; the science just needs the evidence to catch up with the claim.","headline":"A reasonable first SI system for VPI children, but the headline PPMC gains rest on tiny test sets and no significance tests, so they are not yet established as real.","tokens_in":10512,"tokens_out":3022,"would_cite":false,"duration_ms":39449,"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":"A multi-task speech-inversion system estimates nasalance from audio 16.92% more accurately than a prior baseline, and fine-tuning on children with velopharyngeal insufficiency adds another 7.90% improvement.","keywords":["speech inversion","nasalance","velopharyngeal insufficiency","self-supervised learning","multi-task learning","electroglottography","fine-tuning","children's speech"],"falsifier":"Re-run the same experiments with per-speaker bootstrapped confidence intervals for the PPMC differences between baseline and proposed systems; if the intervals overlap at the 95% level on either the adult test set or the VPI-child folds, the claimed relative improvements (16.92% and 7.90%) are not statistically reliable.","tokens_in":9621,"feed_emoji":"🗣️","tokens_out":2787,"duration_ms":33757,"temperature":0.7,"pith_summary":"This paper claims that a speech-inversion system can estimate nasalance (the velopharyngeal port tract variable) directly from acoustic speech, using self-supervised embeddings combined with glottal source features. The authors report a 16.92% relative improvement in correlation over a previous system when tested on healthy adults, and a further 7.90% improvement after fine-tuning on a small set of children with velopharyngeal insufficiency (VPI). If correct, this offers a non-invasive, instrument-free way to monitor VPI in children, replacing uncomfortable procedures like nasopharyngoscopy and nasometry. The key is that multi-task learning with glottal features and variable-length training segments improves accuracy, and that a small amount of in-domain child speech is enough to adapt the model.","feed_headline":"Audio-only model estimates nasalance 16.9% better","feed_subtitle":"Fine-tuning on kids with velopharyngeal insufficiency adds another 7.9%.","key_machinery":"The central mechanism is a multi-task learning framework that stacks representations from all 25 hidden layers of a self-supervised speech model (WavLM-Large) and passes them through bidirectional GRUs to simultaneously predict the VP tract variable (nasalance) and three source features (F0, periodicity, aperiodicity) plus the EGG envelope. The loss is a weighted combination of Pearson correlation and RMSE (α=0.2). Variable-length random segments of 2-5 seconds replace the fixed 2-second segments used previously, improving robustness to duration variation. For children, the EGG-dependent head is removed and the model is fine-tuned on VPI-child speech.","core_discovery":"The paper's central claim is that a multi-task speech-inversion architecture—using WavLM-Large self-supervised embeddings, variable-length training segments (2-5 seconds), and auxiliary prediction of electroglottograph envelope, F0, periodic energy, and aperiodic energy—estimates nasalance with higher Pearson correlation than prior systems. On the healthy-adult test set, this reaches a PPMC of 0.9488, a 16.92% relative improvement over the baseline system of Siriwardena et al. Fine-tuning the adult-trained model on 14 children with VPI (5-fold cross-validation) raises the average PPMC from 0.6357 to 0.6859, a 7.90% relative improvement, and outperforms training from scratch on the child data","pith_inferences":["The paper does not isolate how much of the adult improvement comes from the multi-task glottal features versus the variable-length segment strategy versus the choice of WavLM-Large; a controlled ablation varying these factors independently would sharpen the finding.","Since the EGG signal is unavailable for children, the practical system relies on acoustically derived source features; the relatively small performance drop when EGG is excluded (0.9488 vs. 0.9470 in the adult ablation) suggests EGG may be a convenient but not essential auxiliary input.","The cross-corpus generalization demonstration uses a single sentence; a more varied corpus would test whether the system truly generalizes beyond the training passages or overfits to their phonetic content.","The clinical value would be higher if the system not only tracks nasalance but also reliably distinguishes hypernasality from audible nasal emission—categories the authors mention but do not evaluate separately."],"forward_implications":["If confirmed, this system could enable at-home or remote monitoring of hypernasality and nasal emission in children with VPI, using only a smartphone microphone.","The two-stage training recipe—adult pretraining then fine-tuning on a small clinical dataset—provides a template for adapting speech-inversion models to other disordered or pediatric populations with limited data.","Including glottal source features (EGG, F0, periodicity, aperiodicity) as auxiliary tasks consistently improves nasalance estimation, suggesting vocal-fold dynamics carry complementary information about velopharyngeal function.","The fine-tuned system tracks nasality peaks corresponding to nasal emissions and hypernasal vowels, indicating potential clinical use for locating specific articulatory events in a child's speech.","Because training from scratch on the small VPI-child dataset yields far lower correlation (0.6388), the adult pretraining is shown to be essential, supporting the value of large healthy-speech corpora for clinical adaptation."],"fun_headline_variants":["Audio-only nasalance: 17% better, child fine-tuning adds 8%","Speech inversion gets 17% better nasalance, then 8% more on kids","Noninvasive model estimates nasalance 17% better, then 8% with VPI data","Acoustic-only nasalance model: 17% gain, 8% more after child adaptation","From adults to kids: acoustic model improves nasalance 17% then 8%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported performance improvements are computed on very small test sets—two healthy adult speakers for the adult comparison, and only three children per fold for the VPI fine-tuning—with no significance testing or confidence intervals, so the differences could reflect speaker-specific variation rather than true model improvement.","fun_headline_variants_meta":{"raw":{"variants":["Audio-only nasalance: 17% better, child fine-tuning adds 8%","Speech inversion gets 17% better nasalance, then 8% more on kids","Noninvasive model estimates nasalance 17% better, then 8% with VPI data","Acoustic-only nasalance model: 17% gain, 8% more after child adaptation","From adults to kids: acoustic model improves nasalance 17% then 8%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000468,"raw_usage":{"total_tokens":2153,"prompt_tokens":709,"completion_tokens":1444,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":1329}},"tokens_in":453,"tokens_out":1444,"duration_ms":14923,"temperature":1.0,"reasoning_tokens":1329,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:59:39.927041+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same experiments with per-speaker bootstrapped confidence intervals for the PPMC differences between baseline and proposed systems; if the intervals overlap at the 95% level on either the adult test set or the VPI-child folds, the claimed relative improvements (16.92% and 7.90%) are not statistically reliable.","supporting_citations":[],"review_version":1}