{"id":"c1befd83-a0ce-4f02-8274-53a3755d1331","arxiv_id":"2501.09229","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A greedy tree of linear classifiers and regressors, trained on TitaNet voice embeddings, predicts speaker age on TIMIT with mean absolute error 3.97 years, beating prior published models.","lead":"This paper introduces a piecewise linear model that splits voice features into regions and fits a simple line inside each region to predict a speaker's age. The authors report the lowest average error to date on the TIMIT age-prediction benchmark, though the comparison lacks error bars and code.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on uncontrolled comparison; prior results in Table I may use different splits, features, or augmentation, and no error bars are given.","rationale":"I read the paper in good faith. The algorithm is clearly described and the greedy partitioning is a plausible extension of regression trees. However, the headline result is an empirical SOTA claim, and that claim is only as strong as the comparison. The paper does not specify whether external baselines used the same split, features, or augmentation; it also does not provide statistical uncertainty. The internal baselines (linear regression, random forest, MLP) are presumably on the same features, and TLM beats them, but those are not the 'prior SOTA' referenced in the abstract. The 0.16 margin over Gupta et al. is small relative to likely MAE variability. Thus the central claim needs a controlled comparison and error bars. This matches the reader's weakest assumption; I agree with the CONDITIONAL verdict. No internal contradiction or fatal flaw was found, so the concern is about evidence, not validity. Additionally, the promised code at https://github.com/DareenHarthi/tlm is not yet available, which would enable an independent check. The pseudocode in Algorithm 1 contains a likely typo in the total reduction formula ('errornp − errornl + errornr' instead of subtracting the sum), but the surrounding text clarifies the intended objective. I recommend the same conditional acceptance pending a controlled replication.","tokens_in":8936,"tokens_out":7019,"duration_ms":68698,"concrete_test":"Re-run Gupta et al. [18] under exactly the same protocol used for TLM: same TIMIT speaker-independent train/test split, same 192-d TitaNet-Large embeddings (plus the same two residual blocks and optimization, if any), and same age-conditioned mixup augmentation. If its MAE is at or below 3.97, the claimed new SOTA is an artifact of the comparison protocol. Also compute the standard error of the MAE difference over multiple seeds; if the difference is within one standard error, 'significant margin' is not justified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that TLM achieves MAE 3.97, 'the lowest error ever obtained on these data' (Section V). This claim is supported by Table I, which juxtaposes TLM numbers with previously published results (Singh et al., NRT, Manav et al., Gupta et al.) without evidence that those works used the same TIMIT train/test split, the same TitaNet-Large feature extractor, or the same mixup augmentation. TLM additionally fine-tunes the features via two residual blocks, so the comparison conflates the TLM architecture with a stronger feature representation. The 0.16 MAE margin over Gupta et al. (4.13 vs 3.97) could be entirely due to feature mismatch or protocol differences. Moreover, no confidence intervals or repeated-seed variability are reported, so the margin may not be statistically significant on the 1680-sample test set. The 'significant margin' language is therefore unsupported by the evidence as presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Tessellated Linear Model (TLM), a piecewise-linear regression model that recursively partitions the feature space with binary hyperplane classifiers and fits a linear regressor within each leaf. The method is applied to age prediction from voice on the TIMIT dataset using TitaNet-Large embeddings, and the authors also fine-tune the feature representation with two residual blocks after the tree is built. The central claim is that TLM with feature optimization achieves an MAE of 3.97 on TIMIT, which the paper describes as 'the lowest error ever obtained on these data' and as outperforming all prior reported results.","tokens_in":9150,"tokens_out":4102,"duration_ms":39611,"significance":"If the empirical claim is substantiated, the paper would demonstrate that a simple, interpretable, piecewise-linear model can outperform deep neural networks in a low-data regression setting, which is a meaningful result for voice biometrics and for tabular/small-data regression more broadly. The model design is clearly presented, the oracle baseline is explicitly labeled as such, and the source code is promised. However, the strength of the paper currently rests on an uncontrolled comparison that mixes feature extractors and protocols, and the lack of error bars or significance tests leaves the 'significant margin' unsupported.","major_comments":[{"comment":"The claim that TLM achieves 'the lowest error ever obtained on these data, by a significant margin' rests on a comparison with previously published results (Singh et al. [8], NRT [41], Manav et al. [15], Gupta et al. [18]) without evidence that those works used the same TIMIT train/test split, the same feature extractor, or the same augmentation. Since the TLM pipeline uses TitaNet-Large embeddings and additionally fine-tunes them with two residual blocks, the 0.16 MAE improvement over Gupta et al. (4.13 vs 3.97) could be due to the feature representation rather than the TLM architecture. A controlled comparison in which all baselines are re-run on identical TitaNet features and the same split, with multiple seeds and error bars, is needed to support the SOTA claim.","section":"Table I and Section V"},{"comment":"In the pseudocode, the line 'total reduction ← errornp − errornl + errornr' appears to be a bug: the reduction should be errornp − (errornl + errornr). As written, the criterion does not measure the decrease in squared error, and because 'best reduction' is initialized to −∞, the algorithm would accept any split, including one that increases error. This contradicts the objective stated in Section III-A and affects the correctness of the core greedy splitting procedure.","section":"Algorithm 1"},{"comment":"The statement that 'a single iteration [of alternating optimization] is sufficient' is asserted without sensitivity analysis, and the feature-optimization procedure is underspecified: the two residual blocks are described, but the loss weighting between classifier and regressor terms, the optimizer, learning rate, number of epochs, and the specific value of the maximum tree depth D are not reported. Because feature optimization is a component of the reported method, these details are needed for reproducibility and to assess the contribution of each algorithmic step. Please report results with 0, 1, and 2 iterations of alternating optimization and provide full hyperparameter settings.","section":"Section IV-B"}],"minor_comments":[{"comment":"The word 'polytops' should be 'polytopes'.","section":"Section III"},{"comment":"'V oronoi' should be 'Voronoi'.","section":"Section IV-C"},{"comment":"'signifiantly' should be 'significantly'.","section":"Section V"},{"comment":"The model name is written as 'Titanet-Large' but the reference [24] uses 'TitaNet'; please standardize the spelling.","section":"Section IV-B"},{"comment":"The source code link is provided, but the repository was not accessible at review time; please ensure the code is publicly available upon publication.","section":"Abstract and footnotes"}],"recommendation":"major_revision","confidential_remarks":"The paper's central empirical claim is not yet supported because the SOTA comparison is not controlled, and the Algorithm 1 bug raises concern about the correctness of the greedy criterion. However, these issues are fixable: the authors can re-run baselines under a shared protocol, add error bars and significance tests, correct the pseudocode, and provide the missing implementation details. I do not see a fundamental flaw in the proposed model that would justify rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one. It's a piecewise-linear regression tree with oblique splits and linear leaves, applied to age prediction from voice on TIMIT. The genuinely new bit is the second step: after the tree is built, they freeze it and backprop through the tree's linear classifiers and regressors to fine-tune the TitaNet embeddings. That's a clean idea, and the paper is honest enough to label the oracle row as an oracle.\n\nWhat's not new is the overall architecture—M5 and logistic model trees did oblique splits with linear leaves thirty years ago. The response-threshold split search is a plausible heuristic, not a breakthrough. But the writing is clear, the algorithm is specified in enough detail to reimplement, and the related work covers the relevant lines. No circularity: they train on labeled data and evaluate on the TIMIT test set.\n\nThe problem is the headline claim. Table I juxtaposes TLM numbers with published baselines that used different feature extractors, different train/test splits, and different augmentation. TitaNet plus two residual blocks is a stronger representation than anything the prior baselines had, so the 0.16 MAE margin over Gupta et al. could be entirely from features. There are no error bars, no repeated seeds, and no validation split, so 'significant margin' is rhetorical. The paper promises code but doesn't ship it. These are all fixable in a revision: run the baselines on the same features, same split, same augmentation, and report variance. If the margin survives that, it's a real result.\n\nThe citation pattern is fine; self-citing Singh and Raj is normal when you're extending your own tree work. The oracle row is not a claim, it's a sanity check.\n\nBottom line: this deserves a serious referee, but the referee should require a controlled comparison before accepting the SOTA number. It's a modest incremental paper with one genuinely interesting idea, not a field-changer.","headline":"A modest tree-based method with a nice feature-fine-tuning twist, but the SOTA claim is built on an uncontrolled baseline table and no error bars.","tokens_in":9662,"tokens_out":2729,"would_cite":false,"duration_ms":27580,"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":"The paper claims that a piecewise-linear tessellation of voice features predicts age on TIMIT with mean absolute error 3.97 years, the lowest reported for that dataset, outperforming deep-learning baselines.","keywords":["age estimation from voice","tessellated linear model","piecewise linear regression","regression trees","TIMIT","voice biometrics","feature optimization","small-data regression"],"falsifier":"Run TLM and the previous best method on identical TIMIT train/test folds and identical TitaNet embeddings with the same augmentation, and test whether the MAE gap of 0.16 years persists; if TLM no longer finishes below 4.13, the headline result fails.","tokens_in":8728,"feed_emoji":"🎙️","tokens_out":8190,"duration_ms":80549,"temperature":0.7,"pith_summary":"The paper proposes the Tessellated Linear Model (TLM), a piecewise-linear regressor that splits the voice-feature space into convex regions and fits a separate linear model inside each region. It claims that on the TIMIT corpus, TLM predicts speaker age with a mean absolute error of 3.97 years, the lowest figure reported on that dataset, beating both deep-learning baselines and prior published results by a clear margin. The authors argue this shows a comparatively simple, interpretable model can capture nonlinear voice-age relationships without requiring the large labeled datasets that deep networks need. If the claim holds, it would mean that for small-data biometric regression tasks, greedy, locally linear partitions plus differentiable feature tuning can outperform large neural models.","feed_headline":"Tessellated linear model hits 3.97 MAE on TIMIT age prediction","feed_subtitle":"A simple piecewise-linear regressor with optimized features outperforms published deep-learning baselines on the same corpus.","key_machinery":"The Tessellated Linear Model (TLM) is a piecewise-linear regressor that partitions the feature space $\\mathcal{F}\\subset\\mathbb{R}^d$ into disjoint convex polytope cells and fits an affine model $\\hat{y}(f)=r_\\theta^\\top f+b$ in each cell. The tree is grown greedily: at each node a logistic-regression classifier splits the current data into two sides of a hyperplane, with the split threshold selected by minimizing the summed squared error of the two child regressors. Because every component is linear and differentiable, the feature extractor can then be fine-tuned by backpropagation with the tree frozen. Inference can be hard, routing each input to a single leaf, or soft, blending the predictions along the routing path by classifier probabilities.","core_discovery":"The central discovery is the TLM itself: a hierarchical tessellation in which each cell contains an unconstrained linear regressor, with boundaries and regressors chosen by a greedy recursive search over candidate age thresholds. The paper reports that this model, using 192-dimensional TitaNet speaker embeddings, achieves a mean absolute error of 3.97 and RMSE of 5.36 on TIMIT, below the best prior reported MAE of 4.13. It also reports that even without feature optimization, hard and soft routing versions reach 4.09 and 4.02, respectively, both below the prior best; the gap is attributed to relaxing the continuity-at-boundaries constraint of ReLU networks and to directly optimizing all components for the regression objective. The feature-optimization step freezes the learned tree and updates the embedding network by backpropagation through the differentiable classifiers and regressors.","pith_inferences":["The same greedy partition-plus-backprop recipe should transfer to other small-data regression tasks with differentiable embeddings, though the paper only demonstrates it on voice age.","The oracle experiment (0.49 MAE when true age routes the tree) suggests most residual error is routing error, so better split classifiers could yield larger gains than better leaf regressors.","Iterating tree construction and feature optimization beyond the single pass used here is a natural next test, as is varying the maximum tree depth."],"forward_implications":["If the reported result holds, practitioners with small labeled datasets can get accurate age regression from voice using a tessellated linear model instead of a large deep network.","The feature-optimization step makes the embedding extractor learn from the tree's routing and regression losses, so model and features are adapted to the same objective.","The soft-routing variant, which weights leaf predictions by classifier probabilities, improves MAE from 4.09 to 4.02 over hard routing, showing that smoothing at boundaries helps.","The learned tree partitions the feature space into regions that correspond to age bands, giving a visual and interpretable account of how voice features relate to age."],"supporting_citations":[{"why":"Supplies the TIMIT speech corpus used for all training and testing.","marker":"[5]"},{"why":"Supplies the 192-dimensional TitaNet-Large speaker embeddings used as input features to TLM.","marker":"[24]"},{"why":"Reports the prior best MAE of 4.13 on TIMIT age regression that TLM claims to beat.","marker":"[18]"},{"why":"Provides deep-learning age regression baselines compared in Table I.","marker":"[15]"},{"why":"Provides the Neural Regression Tree baseline for tree-based age prediction.","marker":"[41]"},{"why":"Provides an earlier feature-based age estimation baseline included in the comparison.","marker":"[8]"},{"why":"Supplies the mixup augmentation strategy, modified to mix samples with similar ages.","marker":"[42]"},{"why":"Introduces Regression-via-Classification, a related partition-optimization approach that TLM extends from piecewise constant to piecewise linear.","marker":"[37]"}],"fun_headline_variants":["Piecewise linear model outperforms deep nets on voice age","Tessellated linear model: 3.97 MAE, outperforms deep learning","Voice age prediction: simple tessellated model beats deep net","TLM: piecewise linear regressor beats deep age models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim of beating all prior TIMIT results assumes the earlier published numbers were obtained under the same data split, feature set, and training conditions; if they were not, the 0.16-year margin over the previous best is not a controlled comparison.","fun_headline_variants_meta":{"raw":{"variants":["Piecewise linear model outperforms deep nets on voice age","Tessellated linear model: 3.97 MAE, outperforms deep learning","Voice age prediction: simple tessellated model beats deep net","TLM: piecewise linear regressor beats deep age models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000989,"raw_usage":{"total_tokens":4162,"prompt_tokens":886,"completion_tokens":3276,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":3200}},"tokens_in":502,"tokens_out":3276,"duration_ms":19252,"temperature":1.0,"reasoning_tokens":3200,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:09:01.099754+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TLM and the previous best method on identical TIMIT train/test folds and identical TitaNet embeddings with the same augmentation, and test whether the MAE gap of 0.16 years persists; if TLM no longer finishes below 4.13, the headline result fails.","supporting_citations":[{"cited_title":"Darpa timit acoustic-phonetic continous speech corpus cd-rom. nist speech disc 1-1.1,","cited_arxiv_id":null,"evidence_quote":"Supplies the TIMIT speech corpus used for all training and testing."},{"cited_title":"Titanet: Neural model for speaker representation with 1d depth-wise separable convolutions and global context,","cited_arxiv_id":null,"evidence_quote":"Supplies the 192-dimensional TitaNet-Large speaker embeddings used as input features to TLM."},{"cited_title":"Estimation of speaker age and height from speech signal using bi-encoder transformer mixture model","cited_arxiv_id":"2203.11774","evidence_quote":"Reports the prior best MAE of 4.13 on TIMIT age regression that TLM claims to beat."},{"cited_title":"End-to-end speaker age and height estimation using attention mechanism and triplet loss,","cited_arxiv_id":null,"evidence_quote":"Provides deep-learning age regression baselines compared in Table I."},{"cited_title":"Neural regression trees,","cited_arxiv_id":null,"evidence_quote":"Provides the Neural Regression Tree baseline for tree-based age prediction."},{"cited_title":"Short-term analysis for estimating physical parameters of speakers,","cited_arxiv_id":null,"evidence_quote":"Provides an earlier feature-based age estimation baseline included in the comparison."},{"cited_title":"Regression using classification algorithms,","cited_arxiv_id":null,"evidence_quote":"Introduces Regression-via-Classification, a related partition-optimization approach that TLM extends from piecewise constant to piecewise linear."}],"review_version":1}