{"id":"7f82e015-6314-4958-84c9-58680bd1b39f","arxiv_id":"2508.16176","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A direction-independent latent code learned by an autoencoder lets a small network predict individualized HRTF magnitudes from anthropometric measurements and allows training on multiple datasets.","lead":"This paper proposes a two-stage neural network that predicts a person's head-related transfer functions from head and ear measurements, using a compact latent code that is independent of sound direction. It matters for spatial audio because it offers a way to combine HRTF datasets recorded with different sound-source layouts, which could ease the data shortage in personalized audio.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Autoencoder pretraining may include held-out test subjects, giving the proposed method an unfair advantage over baselines and undermining the central accuracy claim.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the autoencoder may have been pretrained on test subjects, giving the proposed method an unfair advantage. My stress-test reaches the same conclusion from the text of Section 5.1 and Table 2. The concern is concrete and falsifiable: the released code can be inspected and the experiment rerun with a strictly disjoint autoencoder split. If leakage is confirmed, the central claim is unsupported; if not, the paper would still need to address the very small effect size and lack of significance testing. The paper does provide code and the parameter-count advantage is real, but neither mitigates the confound. Therefore the reader's REJECT verdict remains appropriate, and no verdict adjustment is needed.","tokens_in":9691,"tokens_out":5054,"duration_ms":57831,"concrete_test":"Inspect the released repository to determine the exact subject split used for autoencoder pretraining. If test subjects are included in the pretraining set, rerun the full Table 2 comparison with the autoencoder trained exclusively on the 30 CIPIC training subjects (and only the 85 HUTUBS training subjects), explicitly excluding all test subjects and all subjects without anthropometric parameters. Freeze that decoder, retrain both prototype networks, and recompute LSD on the fixed test splits. If the prototype DNN no longer outperforms the HRTF DNN, or the gap becomes statistically insignificant, the reported advantage is a leakage artifact; if the advantage persists with a paired significance test across test subjects, the leakage concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed prototype-based method achieves higher HRTF estimation accuracy than direct DNN baselines (Table 2: 4.97 vs 4.99 dB on CIPIC, 5.01 vs 5.09 dB on HUTUBS). The load-bearing weakness is in the evaluation protocol described in Section 5.1. The paper says the autoencoder 'additionally included 10 and 3 subjects without complete anthropometric parameters for its pretraining' and that 'both the autoencoder and the individualization networks were trained on the same data subsets.' This wording does not state that the held-out test subjects were excluded from autoencoder pretraining; it may even imply that the autoencoder was trained on the full datasets, including test subjects. Since the proposed method uses a frozen decoder to map a predicted prototype to HRTF magnitudes, an autoencoder that saw a test subject's HRTFs during pretraining has subject-specific reconstruction knowledge. The baselines (HRTF DNN and HRTF DM) do not have access to test-subject HRTFs at training time. The reported improvements are only 0.02–0.08 dB, within one standard deviation, so leakage could plausibly explain the entire observed advantage. This makes the comparison in Table 2 insufficiently controlled to support the abstract's claim of high estimation accuracy relative to DNN-based methods.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage HRTF individualization method. First, a source-position- and frequency-conditioned autoencoder is pretrained to map HRTF log-magnitudes to source-position-independent latent vectors ('prototypes') by averaging encoded representations over source positions. Second, a lightweight DNN or a latent diffusion model (DM) is trained to predict these prototypes from anthropometric parameters; the frozen decoder then converts the predicted prototype into an HRTF magnitude. Experiments on CIPIC and HUTUBS compare the proposed prototype DNN/DM with direct HRTF DNN/DM baselines. The proposed prototype DNN reports the lowest LSD in all conditions (e.g., 4.97 vs 4.99 dB on CIPIC; 5.01 vs 5.09 dB on HUTUBS) while using far fewer trainable parameters.","tokens_in":10002,"tokens_out":6170,"duration_ms":63451,"significance":"The core idea of learning a source-position-independent latent space is original and practically relevant, as it enables combining HRTF datasets with different measured source positions without resampling. The manuscript also provides code, which is a strength. If the reported accuracy gains are valid, the method is a useful contribution to HRTF individualization because it drastically reduces the number of parameters and allows multi-dataset training. However, the experimental evidence as presented is not yet convincing: the evaluation protocol is ambiguous on a point that could invalidate the comparison, the parameter-count comparison is confounded, and the small reported differences are not supported by statistical tests.","major_comments":[{"comment":"The evaluation protocol does not unambiguously state whether the held-out test subjects were excluded from autoencoder pretraining. The text says 'both the autoencoder and the individualization networks were trained on the same data subsets' but also says that 10 and 3 subjects 'were additionally included' for autoencoder pretraining. If the autoencoder's pretraining pool included the test subjects (e.g., because it was trained on the full dataset plus the extra subjects), the frozen decoder has already observed each test subject's HRTFs, giving the proposed method access to target information that the baselines do not have. Since the reported LSD differences are 0.02–0.08 dB, well within one standard deviation, leakage could fully explain the observed advantage. This is load-bearing for the central claim. The authors must state explicitly that test subjects were excluded from autoencode","section":"§5.1, Table 2"},{"comment":"The comparison is confounded by the large parameter mismatch (32k vs 82M for the DNN baselines). The baseline HRTF DNN is severely overparameterized relative to the training set size (30–85 subjects), so its performance may be limited by overfitting rather than by the absence of a latent-space intermediate. The proposed method's advantage could stem from the pretrained autoencoder's inductive bias rather than from the source-position-independent prototype itself. To support the claim that the proposed representation is the cause of improvement, the authors should include a baseline with a comparable number of trainable parameters, or a baseline using a linear (e.g., PCA) latent representation.","section":"§5.3, Table 2"},{"comment":"Only five (CIPIC) and six (HUTUBS) test subjects are used. The reported means and standard deviations overlap substantially (e.g., 4.97±0.25 vs 4.99±0.23 dB on CIPIC). Without paired significance testing (e.g., Wilcoxon signed-rank on per-subject LSD differences), the claim that the proposed networks 'consistently outperformed' the baselines is not statistically supported. Please report per-subject results or a significance test.","section":"§5.3, Table 2"}],"minor_comments":[{"comment":"Typo: 'easier to obtained' should be 'easier to obtain'.","section":"§1"},{"comment":"The caption contains stray '🔥 / ❄' symbols, apparently a rendering artifact. Please remove them in the final version.","section":"Fig. 1 caption"},{"comment":"The clamping of the predicted prototype to [−3, 3] at each timestep is ad hoc. Provide a sensitivity analysis or a justification for this range.","section":"§5.2.1"},{"comment":"Panel (d) shows the autoencoder reconstruction, but no quantitative LSD is given for this panel. Adding it would provide a useful reference for the ceiling of the proposed method.","section":"Fig. 3"},{"comment":"The 'N/A' entries for the HRTF DNN are unexplained in the table or its caption; the reason (incompatible source positions across datasets) appears only in §5.3. Move or explain the N/A entries.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the ambiguity in §5.1 about whether the autoencoder was pretrained on the test subjects. If the authors intended the standard interpretation, a one-sentence clarification may suffice; but given the tiny reported margins, the ambiguity is serious enough to require explicit confirmation. If the autoencoder did see test subjects, the central comparison collapses. The parameter-count and significance issues are secondary but should also be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nShort version: the paper's core idea is real — use a source-position-independent latent prototype so an autoencoder can be trained on multiple HRTF databases with different measurement grids, then predict the prototype from anthropometrics and decode. That is a sensible way to attack the small-data problem, and the parameter reduction (32k vs 82M) is striking. Code is released. Worth engaging with.\n\nWhat's new: previous autoencoder personalization methods encode position-dependent latents, so they can't combine CIPIC and HUTUBS directly. The prototype averaging across source positions is a clean solution. The authors also test a diffusion variant, which is a reasonable bonus.\n\nSoft spots: the main evaluation is confounded. Section 5.1 says the autoencoder \"additionally included 10 and 3 subjects without complete anthropometric parameters for its pretraining\" and that \"both the autoencoder and the individualization networks were trained on the same data subsets.\" That doesn't tell the reader whether the held-out test subjects were excluded from autoencoder pretraining. If they were included, the decoder has seen each test subject's HRTF, giving the proposed method an informational advantage the direct baselines don't have. Given the reported gains are 0.02–0.08 dB — well within one standard deviation — the leakage could easily account for the entire observed edge. The paper needs a strictly disjoint autoencoder split and, ideally, a significance test.\n\nA second issue is the baseline set. The comparisons are with two self-built variants of published methods; there is no direct comparison to the autoencoder personalization methods cited in the introduction ([14]–[17]). That leaves the practical contribution less certain.\n\nThe writing is clear and the authors are not overselling — the abstract's \"high estimation accuracy\" is doing more work than the numbers justify. The paper is a good workshop-to-conference quality contribution, but as published it doesn't prove the accuracy claim.\n\nRecommendation: send to peer review, but only with the expectation that the authors either clarify the split (and if leakage exists, redo the experiments) or the paper comes back as a methods-only contribution without the comparative accuracy claim. This deserves referee time, not a desk reject.","headline":"A genuinely neat idea for pooling HRTF datasets with different grids, but the evaluation has a likely leakage flaw that needs to be addressed before the accuracy claim holds.","tokens_in":10472,"tokens_out":3298,"would_cite":false,"duration_ms":34026,"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":"HRTF individualization can be reframed as a two-stage problem: predict a compact source-position-independent latent prototype from a listener's anthropometric measurements, then decode it with a pretrained autoencoder; on CIPIC and HUTUBS t","keywords":["HRTF individualization","anthropometric parameters","latent representation","source-position independence","autoencoder","diffusion model","dataset merging","log-spectral distortion"],"falsifier":"Retrain the autoencoder using only training-set subjects, explicitly excluding the held-out test subjects from pretraining, then rerun the prototype DNN evaluation on the same CIPIC and HUTUBS test sets. If the log-spectral distortion rises to or above the direct-DNN baselines of 4.99 and 5.09 dB, the reported advantage depends on decoder familiarity with test ears rather than on the source-position-independent latent representation.","tokens_in":9572,"feed_emoji":"🎧","tokens_out":6765,"duration_ms":74053,"temperature":0.7,"pith_summary":"The paper proposes making individualized head-related transfer functions by first turning each listener's head, torso, and pinna measurements into a compact latent prototype that does not depend on sound-source direction, then letting a pretrained decoder expand that prototype into full magnitude spectra for every direction. This matters because HRTF datasets are small and use incompatible measurement grids, so direct network training is data-hungry and cannot pool datasets. The paper reports that on CIPIC and HUTUBS, the prototype DNN achieves lower log-spectral distortion than direct DNN baselines (e.g., 4.97 vs 4.99 dB on CIPIC) while using about 32k parameters instead of 82M, and that the latent-space diffusion model also generalizes better across datasets than an HRTF-space diffusion model.","feed_headline":"32k-parameter latent prototype beats 82M-parameter HRTF DNN","feed_subtitle":"On CIPIC and HUTUBS, the two-stage predictor cuts spectral error and merges incompatible direction grids.","key_machinery":"The central object is a source-position-conditioned and frequency-conditioned autoencoder whose encoder produces per-source-position latent codes that are averaged over source positions into a 'prototype.' The prototype is source-position-independent, so the decoder can be conditioned on any direction at synthesis time, and the architecture can train on datasets with incompatible measurement grids. The prototype is the bottleneck through which anthropometry is mapped to HRTF: a small DNN or a latent diffusion model learns to predict prototypes from anthropometric parameters, and the pretrained decoder turns them into HRTF magnitudes.","core_discovery":"The central discovery is that HRTF personalization can be decomposed: a subject-specific latent 'prototype' independent of source position captures who the listener is, and direction/frequency structure is delegated to a conditioned autoencoder decoder. Once the decoder is pretrained, estimating an HRTF for a new listener reduces to predicting a low-dimensional prototype from 23 anthropometric parameters; the same frozen decoder then synthesizes all directions. On CIPIC and HUTUBS test sets, this two-stage scheme outperforms DNNs and diffusion models that estimate HRTF magnitudes directly, and it is the only scheme in the comparison that can be jointly trained on both datasets because the la","pith_inferences":["The reported accuracy margins over direct DNN baselines are small (roughly 0.02–0.12 dB); the larger practical win may be parameter efficiency and the ability to merge datasets, not dramatically lower spectral error.","If the paper's pretraining protocol does not exclude test subjects from the autoencoder's pretraining pool, as Section 5.1's wording leaves open, the frozen decoder may already be familiar with the test subjects' ears; retraining with strict subject exclusion would settle whether the accuracy gain survives.","A source-position-independent prototype could serve as a reusable subject embedding for other spatial-audio tasks, such as similarity-based HRTF retrieval, transfer to unmeasured directions, or uncertainty-aware probabilistic rendering when generated by a diffusion model.","The same two-stage decomposition may transfer to other low-data, high-dimensional personalization problems where global structure is shared across conditions but subject-specific variation is low-dimensional."],"forward_implications":["Multiple HRTF datasets with different measured direction sets can be pooled in one training run, directly attacking the small-data problem in HRTF individualization.","Individualizing a new listener becomes a lightweight regression task: predict a 64-dimensional prototype per frequency from 23 anthropometric measurements instead of the full magnitude grid.","The direct 82M-parameter HRTF DNN can be replaced by a roughly 32k-parameter prototype DNN plus a shared decoder, with lower log-spectral distortion on the tested datasets.","Latent-space diffusion models inherit the same dataset-pooling benefit, whereas HRTF-space diffusion cannot be jointly trained across CIPIC and HUTUBS because of incompatible source positions.","Because the decoder is frozen after pretraining, the per-subject cost of deployment scales only with the prototype estimation network."],"supporting_citations":[{"why":"Supplies the autoencoder with source-position conditioning that the proposed latent-prototype method adapts.","marker":"[18]"},{"why":"Extends the autoencoder with frequency conditioning and Fourier feature mapping, providing the architecture details used here.","marker":"[19]"},{"why":"The CIPIC dataset provides measured HRTFs, anthropometric parameters, and the train/test split for one evaluation setting.","marker":"[20]"},{"why":"The HUTUBS dataset provides a second HRTF database with different source positions, enabling the cross-dataset pooling test.","marker":"[21]"},{"why":"A DNN that directly estimates HRTFs from anthropometric parameters; serves as the main non-diffusion baseline.","marker":"[7]"},{"why":"A denoising-diffusion approach operating directly in HRTF space; serves as the diffusion baseline.","marker":"[30]"},{"why":"The latent diffusion model framework on which the prototype DM is built.","marker":"[34]"},{"why":"Provides the DDIM sampler used for deterministic accelerated denoising of prototypes.","marker":"[36]"},{"why":"Classifier-free guidance is used during inference to strengthen conditioning on anthropometric features.","marker":"[37]"}],"fun_headline_variants":["HRTF personalization via shared latent, not direct magnitude","Two-stage HRTF: 23 anthropometrics to latent, then decoder","32k-parameter HRTF model outperforms 82M-parameter DNN","Combine CIPIC and HUTUBS HRTF sets with one decoder","Position-independent prototype cuts HRTF estimation cost"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The accuracy comparison assumes that pretraining the decoder on some of the same subjects later used for testing does not leak their ear characteristics; the paper adds subjects without anthropometric data for pretraining (Section 5.1) and never states that test subjects are excluded from that pretraining pool.","fun_headline_variants_meta":{"raw":{"variants":["HRTF personalization via shared latent, not direct magnitude","Two-stage HRTF: 23 anthropometrics to latent, then decoder","32k-parameter HRTF model outperforms 82M-parameter DNN","Combine CIPIC and HUTUBS HRTF sets with one decoder","Position-independent prototype cuts HRTF estimation cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000201,"raw_usage":{"total_tokens":1176,"prompt_tokens":667,"completion_tokens":509,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":411,"completion_tokens_details":{"reasoning_tokens":420}},"tokens_in":411,"tokens_out":509,"duration_ms":5906,"temperature":1.0,"reasoning_tokens":420,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:28:42.970338+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the autoencoder using only training-set subjects, explicitly excluding the held-out test subjects from pretraining, then rerun the prototype DNN evaluation on the same CIPIC and HUTUBS test sets. If the log-spectral distortion rises to or above the direct-DNN baselines of 4.99 and 5.09 dB, the reported advantage depends on decoder familiarity with test ears rather than on the source-position-independent latent representation.","supporting_citations":[{"cited_title":"Head-related transfer function interpolation from spatially sparse measurements using autoencoder with source position conditioning,","cited_arxiv_id":null,"evidence_quote":"Supplies the autoencoder with source-position conditioning that the proposed latent-prototype method adapts."},{"cited_title":"Spatial upsampling of head-related transfer function using neural network conditioned on source position and frequency,","cited_arxiv_id":null,"evidence_quote":"Extends the autoencoder with frequency conditioning and Fourier feature mapping, providing the architecture details used here."},{"cited_title":"The CIPIC HRTF database,","cited_arxiv_id":null,"evidence_quote":"The CIPIC dataset provides measured HRTFs, anthropometric parameters, and the train/test split for one evaluation setting."},{"cited_title":"Brinkmann, D","cited_arxiv_id":null,"evidence_quote":"The HUTUBS dataset provides a second HRTF database with different source positions, enabling the cross-dataset pooling test."},{"cited_title":"Prediction of head related transfer functions using machine learning approaches,","cited_arxiv_id":null,"evidence_quote":"A DNN that directly estimates HRTFs from anthropometric parameters; serves as the main non-diffusion baseline."},{"cited_title":"Towards HRTF personalization using denoising diffusion models,","cited_arxiv_id":null,"evidence_quote":"A denoising-diffusion approach operating directly in HRTF space; serves as the diffusion baseline."},{"cited_title":"High-resolution image synthesis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"The latent diffusion model framework on which the prototype DM is built."},{"cited_title":"Denoising diffusion implicit models,","cited_arxiv_id":null,"evidence_quote":"Provides the DDIM sampler used for deterministic accelerated denoising of prototypes."}],"review_version":1}