{"id":"05fe157a-3194-47f6-9198-cf270bf85bfe","arxiv_id":"2506.18954","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SHAMaNS localizes multiple sound sources by fusing an alpha-stable spatial measure with a neural steering vector interpolator, outperforming MUSIC and SRP-PHAT in simulated scenes.","lead":"This paper combines an alpha-stable statistical model for sound localization with a neural network that fills in missing steering vectors from sparse measurements. The hybrid method reports lower localization error than MUSIC and SRP-PHAT in simulated multi-source scenes, using only a fraction of the measured directions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The derivation in Sec. III-A is incorrect: adding C_alpha 1_L to Eq. (10) does not leave the beta-divergence update unchanged, because the constant appears in the gradient; the claimed noise-immunity of the algorithm is therefore unsupported.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern as I do, and it is the most consequential issue in the paper. The central quantitative comparisons in Figs. 4 and 5 are implemented with Eq. (5), so the arithmetic error does not unwind the reported scores. Nevertheless, the paper's reasoning for fusing Neural Steerer with alpha-stable SSL rests on the assertion that additive NS error is implicitly modeled; that assertion is the paper's main theoretical contribution. If it is false, the method is a heuristic combination whose advantage over SRP-PHAT and MUSIC is not supported by a principle. I also considered the absence of the prior alpha-stable SSL method [25] and the original Neural Steerer [29] as baselines; that is a real secondary limitation, but it is less squarely load-bearing than the explicit mathematical claim in Sec. III-A because the strongest claim as stated concerns the comparison against SRP-PHAT and MUSIC. Credit is due for releasing the code and reporting concrete experimental settings, which aids reproducibility, but no formal verification backs the derivation. The verdict should remain CONDITIONAL: either the derivation must be corrected or the noise-robustness claim withdrawn, while the empirical evaluation may still stand as an engineering result.","tokens_in":8815,"tokens_out":8703,"duration_ms":87327,"concrete_test":"Derive the beta-divergence fixed-point update for the objective in Eq. (10) symbolically, or numerically for a scalar case: set L=1, Psi=1, I=1.5, C_alpha=0.5, lambda→0. The full-model beta=2 objective (I - (Upsilon + C_alpha))^2 has fixed point Upsilon = I - C_alpha = 1.0, while Eq. (5) with beta=2 has fixed point Upsilon = I = 1.5. If this mismatch reproduces, Eq. (5) is not the update corresponding to Eq. (10), confirming that the claim of Section III-A is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A claims that introducing additive isotropic alpha-stable noise n_ft yields the Levy exponent If = Psi_f Upsilon + C_alpha 1_L (Eq. 10), and that deriving the multiplicative update along Upsilon makes C_alpha 1_L vanish, leaving Eq. (5). This is not correct. For beta-divergence D_beta(I | y) with y = Psi Upsilon + C_alpha 1_L, the gradient with respect to Upsilon is Psi^T [ y^{beta-2} ⊙ (y - I) ] (up to convention-dependent prefactors). The constant C_alpha 1_L appears explicitly in y and does not cancel. For beta = 1, the KL gradient is Psi^T(1_L - I ⊘ (Psi Upsilon + C_alpha 1_L)); the denominator contains C_alpha 1_L. For beta = 2, the gradient is Psi^T(Psi Upsilon + C_alpha 1_L - I). Only if the noise term is absent, or treated as part of the data as I' = I - C_alpha 1_L, would Eq. (5) follow, and neither is stated. Consequently, the theoretical mechanism claimed for SHAMaNS — that estimating alpha makes the algorithm automatically robust to the Neural Steerer reconstruction error absorbed in n_ft — is unsupported. The code and experiments use Eq. (5), so the empirical localization results are not invalidated by this error; however, the paper's core theoretical justification for the hybrid noise model is broken as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SHAMaNS, a sound source localization method that combines an alpha-stable spatial measure model with a physics-informed neural network, Neural Steerer, for interpolating measured steering vectors. The authors extend the alpha-stable mixing model with an additive isotropic alpha-stable noise term, claiming that the noise term vanishes from the multiplicative update for the spatial measure, thereby making the algorithm automatically robust to interpolation errors. The method is evaluated on simulated multi-source scenarios based on the SPEAR dataset, comparing against MUSIC and SRP-PHAT with oracle, algebraic, and Neural Steerer interpolated steering vectors. The reported results indicate improved localization accuracy for multiple concurrent sources and near-oracle performance with a small number of measured steering vectors.","tokens_in":9193,"tokens_out":2822,"duration_ms":31551,"significance":"If the central claims hold, SHAMaNS would be a practically useful contribution: it targets a realistic setup where steering vectors are measured at sparse directions, and it releases code and uses a publicly available dataset, which supports reproducibility. The idea of modeling Neural Steerer reconstruction error as impulsive noise is conceptually appealing, and the empirical setting is relevant to augmented-reality and hearable applications. However, the main theoretical justification in Section III-A is not supported by the derivation as written, and the evaluation lacks the natural ablations that would isolate the contribution of each component. The paper is therefore significant in scope but currently requires substantial revision before its claims can be accepted.","major_comments":[{"comment":"The derivation that the additive noise term C_alpha 1_L vanishes from the multiplicative update is incorrect. For the model I_f = Psi_f Upsilon + C_alpha 1_L, the beta-divergence objective depends on Upsilon through the sum Psi Upsilon + C_alpha 1_L, and the gradient with respect to Upsilon contains this sum explicitly. For example, for beta=1 the gradient is Psi^T(1_L - I ⊘ (Psi Upsilon + C_alpha 1_L)), and for beta=2 it is Psi^T(Psi Upsilon + C_alpha 1_L - I). The constant does not cancel in either case. Equation (5) follows only if the noise term is absent or if the observed data are replaced by I' = I - C_alpha 1_L, which is not stated. Consequently, the claim that a correct estimate of alpha makes the algorithm implicitly robust to the Neural Steerer reconstruction error is unsupported by the derivation. This is a load-bearing issue because the paper's theoretical justification for the hybrid model rests on this step.","section":"Section III-A, Eqs. (9)-(10) and Eq. (5)"},{"comment":"The empirical evaluation omits the two most natural baselines for the proposed fusion: the alpha-stable SSL method without Neural Steerer (reference [25]) and Neural Steerer without the alpha-stable spatial measure (reference [29]). Since the paper's contribution is precisely the fusion of these two components, the reported comparisons against MUSIC and SRP-PHAT do not by themselves demonstrate that the hybrid combination is beneficial. The authors should include these ablations, or at least provide a clear explanation of why they are not applicable, before claiming that SHAMaNS outperforms state-of-the-art methods.","section":"Section IV-B, Figs. 4 and 5"},{"comment":"The main multi-source accuracy and AUC results are reported without error bars, confidence intervals, or statistical significance tests, even though the experiments use only 30 random scenes per condition. The text states that SHAMaNS outperforms baselines for 2 to 6 sources and that accuracy exceeds 55% for N<5, but without variance estimates it is not possible to assess whether these differences are meaningful. The authors should provide per-condition standard deviations or otherwise quantify the uncertainty of the reported curves.","section":"Section IV-B, Figs. 4 and 5"}],"minor_comments":[{"comment":"The caption states \"RT60=0.123 ms,\" which appears to be a unit error; the text elsewhere uses seconds, and 0.123 ms would be an unrealistically short reverberation time. Please correct the unit, likely to seconds or milliseconds as appropriate.","section":"Fig. 5 caption"},{"comment":"The statement that \"oracle SSL performance could be obtained with only 10% of the initial measurements\" is imprecise: with NSV=128 out of 1020 measured steering vectors, the fraction is about 12.5%, not 10%. Please clarify the exact number of measurements used to support this claim.","section":"Section IV-A"},{"comment":"The estimator in Eq. (3) is written with an unusual combination of absolute value and logarithm after the empirical mean; please verify that the formula matches the standard estimator from [25], and add a reference or a derivation for this exact expression.","section":"Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"The paper heavily builds on the authors' prior work in references [19], [25], and [29]. Given that the novelty is the fusion of the alpha-stable SSL method and the Neural Steerer, the absence of these two components as baselines in the experimental section is particularly concerning and should be addressed in revision. The theoretical flaw in Section III-A is also central to the paper's framing; if it cannot be corrected, the authors should substantially reframe their contribution as an empirical combination rather than a theoretically motivated noise-robust model."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the empirical system is worth a look, but the paper's central theoretical claim—that adding isotropic alpha-stable noise leaves the multiplicative update unchanged—is simply wrong, and the evaluation misses the two most natural baselines. The combination of Neural Steerer with alpha-stable spatial measure localization is new, and the results for 2-6 sources on measured smartglasses SVs are genuinely interesting. The NS modification to predict SH coefficients is a reasonable engineering choice, and the observation normalization is sensible. The paper also ships code, which is more than many.\n\nThe problem is Section III-A. If the Levy exponent is Psi Upsilon + C_alpha 1_L, then the beta-divergence gradient with respect to Upsilon contains that constant in the denominator/update. It does not vanish. For beta=1, the KL gradient is Psi^T(1 - I ⊘ (Psi Upsilon + C_alpha 1_L)); the constant is right there. So Eq. (5) is not the update for the stated model. The only way to get Eq. (5) is to treat the noise as part of the data or drop it, which defeats the purpose. This is not a nitpick: the paper claims that estimating alpha automatically absorbs Neural Steerer reconstruction error, and that claim is unsupported. The algorithm still works empirically, but the mechanism is not what the paper says.\n\nThe evaluation has two soft spots. First, the natural baselines are missing: the alpha-stable SSL of [19,25] and the original Neural Steerer [29]. Without them, the gains from the hybrid fusion can't be attributed. Second, Figs. 4 and 5 show no variance or error bars, and with only 30 scenes per condition, the differences could be within noise. The N=1 results in Fig. 2 do have standard deviations, so this is fixable.\n\nThe self-citation pattern is fine; the cited prior work is external and relevant. The code availability is a plus.\n\nBottom line: this is a promising engineering paper with a load-bearing theoretical error in the writeup. The empirical contribution deserves peer review, but the noise-model derivation must be corrected or removed, and the missing baselines added. A serious referee would catch this, and it should not be desk-rejected.","headline":"Promising system, but the advertised noise-robustness theory does not survive contact with the gradient.","tokens_in":9758,"tokens_out":1977,"would_cite":false,"duration_ms":20297,"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":"SHAMaNS claims that an alpha-stable spatial measure fed by neural-interpolated steering vectors outperforms SRP-PHAT and MUSIC for two to six concurrent sources, using about ten percent of the measured steering vectors.","keywords":["sound source localization","alpha-stable distributions","spatial measure","steering vector interpolation","Neural Steerer","physics-informed deep learning","multi-source localization","microphone arrays"],"falsifier":"Compute the gradient of the $\\beta$-divergence between $\\hat{I}_f$ and $\\Psi_f \\Upsilon + C_\\alpha \\mathbf{1}_L$ with respect to $\\Upsilon$; if the constant does not cancel, run a synthetic test with known $\\Upsilon$, elliptically contoured noise, and increasing noise level, and check whether SHAMaNS's estimated spatial measure degrades as the noise grows. If accuracy is independent of the noise level, the absorption claim survives; if it drops, the update in Eq. (5) is not justified by the stated model.","tokens_in":8638,"feed_emoji":"🎙️","tokens_out":10021,"duration_ms":91194,"temperature":0.7,"pith_summary":"The paper proposes SHAMaNS, a sound source localization method that fuses an $\\alpha$-stable statistical model of the multichannel mixture with a coordinate-based neural network, Neural Steerer, that interpolates measured steering vectors to directions where no measurement was taken. The paper argues the fusion is theoretically coherent because the interpolator's reconstruction error behaves like impulsive, heavy-tailed noise and is absorbed by the $\\alpha$-stable noise term. It reports that SHAMaNS outperforms SRP-PHAT and MUSIC when locating two to six concurrent sources, keeps accuracy above 55 percent for fewer than five sources, and reaches oracle-level performance using only about 10 percent of the measured steering vectors for the neural interpolator. If right, the method would make high-resolution multi-source localization practical on small wearable arrays with sparse acoustic calibration.","feed_headline":"Ten percent of steering vectors matches full-array localization","feed_subtitle":"Alpha-stable spatial measure plus neural interpolation beats SRP-PHAT and MUSIC for two to six concurrent sources.","key_machinery":"The load-bearing object is the $\\alpha$-stable spatial measure $\\Upsilon = [\\Upsilon_1,\\dots,\\Upsilon_L] \\in \\mathbb{R}_+^L$, a non-negative vector over candidate source positions whose nonzero entries mark active sources; $\\alpha$-stable distributions are heavy-tailed laws with characteristic exponent $\\alpha \\in (0,2]$ that yield a unique spatial measure when $\\alpha<2$. The measure is estimated through the Lévy exponent relation $\\hat{I}_f \\approx \\Psi_f \\Upsilon$, where $\\Psi_f$ contains the normalized inner products $|\\tilde{a}_{lf}^H \\tilde{a}_{l'f}|^\\alpha$ between steering vectors, stacked over frequencies as $\\hat{I}\\approx \\Psi\\Upsilon$, and recovered by minimizing the $\\beta$-divergence with a sparsity penalty, producing the multiplicative update of Eq. (5). The second piece is Neural Steerer, a coordinate-based network that outputs spherical-harmonic expansion coefficients $c_{\\nu\\mu,mf}$ for each microphone, direction, and frequency, so a few hundred measured steering vectors can be upsampled to the 60 candidate directions. Observations are normalized as $\\tilde{x}_{ft}=x_{ft}/\\|x_{ft}\\|_p^p$ with $p<\\alpha$, which the paper argues leaves the spatial measure unchanged up to a direction-independent scale while taming outliers.","core_discovery":"On the paper's own terms, the central discovery is that the $\\alpha$-stable spatial measure remains a valid localization statistic even when the steering vectors entering it come from a neural interpolator rather than from measurements at every candidate direction. The paper extends the mixing model with an additive noise term $n_{ft}$ modeled by an elliptically contoured isotropic $\\alpha$-stable distribution, and claims that the constant noise contribution $C_\\alpha \\mathbf{1}_L$ vanishes from the multiplicative update for the spatial measure, so the noiseless update applies and the Neural Steerer's frequency-local interpolation error is handled implicitly. The experimental claim is that with 32 measured steering vectors (about 10 percent of the available 1020) the interpolated steering vectors match the localization accuracy obtained with oracle measured steering vectors, and that SHAMaNS beats SRP-PHAT and MUSIC for two to six simultaneous sources.","pith_inferences":["The same absorption mechanism should transfer to other learned steering-vector or transfer-function models whose residuals are concentrated in certain frequencies, such as HRTF upsamplers, even without the Neural Steerer architecture.","Because the normalized observations keep the spatial measure unchanged up to scale, the pipeline is a natural candidate for arrays with unknown per-microphone gain calibration; that is a testable stress case the paper does not run.","A direct extension would estimate $\\alpha$ jointly with the spatial measure inside the optimization rather than fixing it in advance, which would show whether the robustness comes from the heavy-tailed model or mainly from the interpolation quality."],"forward_implications":["With roughly 32 randomly sampled measured steering vectors, SHAMaNS matches the localization accuracy obtained with the full set of 1020 oracle steering vectors.","For two to six concurrent sources at 20 dB SNR and low-to-moderate reverberation, SHAMaNS reports higher accuracy than SRP-PHAT and MUSIC, with accuracy above 55 percent for fewer than five sources.","The $\\alpha$-stable model treats neural steering-vector interpolation error as impulsive noise, so the pipeline needs no separate uncertainty calibration of the interpolator.","All methods degrade when algebraic steering vectors are used, meaning the practical gain of SHAMaNS depends on the quality of the interpolated steering vectors, and the paper argues a correct estimate of $\\alpha$ is what makes the noise absorption work."],"supporting_citations":[{"why":"Supplies the alpha-stable spatial measure formulation for heavy-tailed source localization and the beta-divergence sparsity-penalized estimator that SHAMaNS inherits.","marker":"[19]"},{"why":"Establishes the unique discrete spatial measure for multichannel alpha-stable mixtures and the Lévy exponent estimation used in Eq. (3).","marker":"[25]"},{"why":"Introduces the Neural Steerer architecture that this paper adapts to synthesize steering vectors from sparse measurements.","marker":"[29]"},{"why":"Provides the elliptically contoured alpha-stable noise model that SHAMaNS uses to absorb Neural Steerer reconstruction error.","marker":"[22]"},{"why":"Supplies the SPEAR head-worn array and its measured steering vectors used to train the interpolator and evaluate localization.","marker":"[34]"},{"why":"MUSIC is the eigendecomposition baseline against which multi-source localization accuracy is compared.","marker":"[7]"},{"why":"SRP-PHAT is the beamforming baseline against which multi-source localization accuracy is compared.","marker":"[8]"},{"why":"Alpha-stable distribution theory, including the covariation formula used to justify the observation normalization.","marker":"[30]"},{"why":"Supplies the method used to estimate the characteristic exponent alpha from the observed mixture before localization.","marker":"[36]"}],"fun_headline_variants":["10% of steering vectors, full localization accuracy","Neural steerer cuts steering vectors to 10%, maintains accuracy","Alpha-stable measure with neural interpolation beats MUSIC","Sound localization from 10% of steering vectors","Multiple sources localized with sparse steering vectors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that when the $\\alpha$-stable model is extended with additive noise, the noise term $C_\\alpha \\mathbf{1}_L$ disappears from the multiplicative update, leaving the noiseless update valid without separately modeling the interpolator's error; if that term actually survives in the gradient, the stated derivation no longer supports the algorithm.","fun_headline_variants_meta":{"raw":{"variants":["10% of steering vectors, full localization accuracy","Neural steerer cuts steering vectors to 10%, maintains accuracy","Alpha-stable measure with neural interpolation beats MUSIC","Sound localization from 10% of steering vectors","Multiple sources localized with sparse steering vectors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000468,"raw_usage":{"total_tokens":2302,"prompt_tokens":882,"completion_tokens":1420,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":1346}},"tokens_in":498,"tokens_out":1420,"duration_ms":12617,"temperature":1.0,"reasoning_tokens":1346,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:45:46.198204+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the gradient of the $\\beta$-divergence between $\\hat{I}_f$ and $\\Psi_f \\Upsilon + C_\\alpha \\mathbf{1}_L$ with respect to $\\Upsilon$; if the constant does not cancel, run a synthetic test with known $\\Upsilon$, elliptically contoured noise, and increasing noise level, and check whether SHAMaNS's estimated spatial measure degrades as the noise grows. If accuracy is independent of the noise level, the absorption claim survives; if it drops, the update in Eq. (5) is not justified by the stated model.","supporting_citations":[{"cited_title":"Sketching for nearfield acoustic imaging of heavy-tailed sources,","cited_arxiv_id":null,"evidence_quote":"Supplies the alpha-stable spatial measure formulation for heavy-tailed source localization and the beta-divergence sparsity-penalized estimator that SHAMaNS inherits."},{"cited_title":"Scalable source localization with multichannel α-stable distributions,","cited_arxiv_id":null,"evidence_quote":"Establishes the unique discrete spatial measure for multichannel alpha-stable mixtures and the Lévy exponent estimation used in Eq. (3)."},{"cited_title":"Neural steerer: Novel steering vector synthesis with a causal neural field over frequency and direction,","cited_arxiv_id":null,"evidence_quote":"Introduces the Neural Steerer architecture that this paper adapts to synthesize steering vectors from sparse measurements."},{"cited_title":"Elliptically contoured alpha-stable representation for MUSIC- based sound source localization,","cited_arxiv_id":null,"evidence_quote":"Provides the elliptically contoured alpha-stable noise model that SHAMaNS uses to absorb Neural Steerer reconstruction error."},{"cited_title":"An introduction to the speech enhancement for augmented reality (SPEAR) challenge,","cited_arxiv_id":null,"evidence_quote":"Supplies the SPEAR head-worn array and its measured steering vectors used to train the interpolator and evaluate localization."},{"cited_title":"Multiple emitter location and signal parameter estimation,","cited_arxiv_id":null,"evidence_quote":"MUSIC is the eigendecomposition baseline against which multi-source localization accuracy is compared."},{"cited_title":"Robust local- ization in reverberant rooms,","cited_arxiv_id":null,"evidence_quote":"SRP-PHAT is the beamforming baseline against which multi-source localization accuracy is compared."},{"cited_title":"Samorodnitsky and M","cited_arxiv_id":null,"evidence_quote":"Alpha-stable distribution theory, including the covariation formula used to justify the observation normalization."},{"cited_title":"Alpha-stable autoregressive fast multichannel nonnegative matrix fac- torization for joint speech enhancement and dereverberation","cited_arxiv_id":null,"evidence_quote":"Supplies the method used to estimate the characteristic exponent alpha from the observed mixture before localization."}],"review_version":2}