{"id":"c9a1a707-53b7-4eb2-b437-9dd6762753b2","arxiv_id":"2505.20961","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A masked-autoencoder SSL model with sparse cross-attention and pretrained audio embeddings reports the best localization error on LuViRA music3 and speech3 while also estimating faulty microphone positions.","lead":"This paper proposes a neural network that locates 3D sound sources from microphone recordings while also estimating the positions of faulty or unknown microphones. It might interest engineers who need sound localization with fewer, cheaper, or less reliably positioned microphones.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (8) defines the source-localization loss over source pairs; for the single-source case K=1 the sum is empty, so Tables 1 and 2 have no stated training signal for source position prediction.","rationale":"I read the paper in good faith. The authors propose a plausible pipeline: pretrained audio embeddings, position embeddings, TDOA features with adaptive coherence weighting, sparse cross-attention, and a masked autoencoder objective. The efficiency claims and the use of pretrained BEATs are reasonable ideas, and the multi-source pairwise loss makes sense as a permutation-invariant objective for K>1. However, the single-source experiments dominate the evaluation, and the stated loss in Eq. (8) is empty when K=1. The masked audio reconstruction and microphone-coordinate losses do not supervise the source position explicitly, because the predicted source position is not fed back into either loss. Therefore the strongest reported results—Table 1 and Table 2—lack a stated training signal. This is not a matter of disagreeing with current practice; it is an internal inconsistency between the method description and the experiments. The reader's weakest-assumption analysis identifies exactly this issue, and I agree. A corrected loss equation or a released training loop would resolve the concern, but neither is present. I also note that the fault-tolerance claim is narrower than the abstract suggests: Table 3 evaluates microphone-position estimation, not source-localization accuracy under faulty or unknown microphones. That is secondary to the Eq. (8) problem but reinforces the need for code or additional experiments. Given the load-bearing inconsistency, the reader's REJECT verdict is appropriate, and I do not adjust it.","tokens_in":8284,"tokens_out":3035,"duration_ms":35562,"concrete_test":"Obtain the released training code and inspect the implementation of Eq. (8); if the source-localization term is literally a pair-distance loss, run a single-source (K=1) training step and measure the gradient norm flowing into the position decoder's output. If that gradient is zero, or if no absolute-coordinate term exists anywhere in the objective, then the single-source results in Tables 1 and 2 cannot arise from the stated training objective. Then retrain with the corrected absolute-coordinate loss and compare the resulting MAE values against Tables 1 and 2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central results are single-source localization (Tables 1 and 2), but the only source-localization term in Eq. (8) is a sum over source pairs: λ_s-loc Σ_{(i,j)∈K} | ||r̂_i - r̂_j||² - ||r_i - r_j||² |. For K=1 this sum is empty, so the position decoder receives no gradient from any source-coordinate objective. The first term of Eq. (8) supervises reconstructed audio, and the second supervises microphone positions; neither term is shown to determine the predicted source location. The text before Eq. (8) says the objective combines reconstruction, microphone-coordinate, and source-localization losses, but no absolute-coordinate loss is stated anywhere. Thus the reported single-source MAE and accuracy values cannot be produced by the objective as written. This is an internal inconsistency in the central claim, not a disagreement with current consensus, and it is fixable by a corrected loss equation or by code release, but the preprint as written does not establish its main results.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 3D sound source localization framework that combines a pretrained audio encoder (BEATs), NGCC-PHAT features with adaptive signal coherence metrics, sparse cross-attention, and a masked autoencoder-style reconstruction. It claims state-of-the-art single-source localization on the LuViRA music3 and speech3 datasets, improved performance with fewer microphones, and the ability to estimate faulty microphone positions. The training objective is stated as a weighted sum of reconstruction loss, microphone-coordinate loss, and a pair-wise source localization loss. The experiments compare against multilateration, DI-NN, GNN, and wav2pos baselines.","tokens_in":8504,"tokens_out":3526,"duration_ms":37418,"significance":"If the results hold, the framework would offer a meaningful improvement in 3D sound source localization efficiency and robustness, particularly the capability to operate with fewer microphones and to estimate faulty microphone positions. The main strengths are the empirical comparisons on the public LuViRA dataset, the reported gains in MAE and acc@30cm, and the ablation-style study of microphone reduction and fault tolerance. However, the central claims rest on a training objective that is internally inconsistent for the single-source experiments that constitute the main results, and the fault-tolerance evaluation does not directly measure source localization error under microphone faults. The paper also omits key implementation details and does not release code, limiting reproducibility. The contribution is potentially valuable but not currently established by the manuscript as written.","major_comments":[{"comment":"The source localization loss in Eq. (8) is defined as a sum over source pairs (i,j)∈K of the absolute difference of squared pairwise distances. For K=1, which is exactly the setting of Tables 1 and 2, this sum is empty, so the position decoder receives no gradient from any source-coordinate objective. The text preceding Eq. (8) states that the loss combines reconstruction, microphone-coordinate, and source-localization losses, but no absolute-coordinate loss is given anywhere. The reported single-source results therefore cannot be produced by the objective as written. This is a load-bearing internal inconsistency in the central empirical claim; the authors must either state the correct loss (e.g., an L1/L2 loss on source coordinates) or release the training code to demonstrate the actual objective used.","section":"Section 3.2.3, Eq. (8)"},{"comment":"The fault-tolerance claims in the abstract and introduction are supported only by microphone-position MAE, not by source localization error under faulty or unknown microphone positions. Table 3 reports MAE for estimating unknown microphone locations in scenes A and B, but the paper never reports source localization MAE in those fault scenarios. Since the claimed benefit is that the system remains accurate for source localization when microphones fail or have unknown positions, the evidence presented is indirect and incomplete. The authors should report source localization error under faulty-microphone conditions (both with known and estimated microphone positions) to substantiate the fault-tolerance claim.","section":"Section 4.3, Table 3"},{"comment":"Reproducibility is severely limited by missing hyperparameters and design choices that are central to the method. The ASCM coherence exponent α in Eq. (4), the number of learned NGCC-PHAT filters P, the sparse cross-attention top-T, the masking ratio used in training, and the loss weights λ_sound, λ_m-loc, and λ_s-loc are not specified. Furthermore, no code or trained models are provided, and the baseline comparison relies on 'following their original hyperparameters' without reporting whether the same data splits, evaluation protocol, and training budgets were used. Without these details, the benchmark comparisons in Tables 1 and 2 cannot be independently verified or reproduced.","section":"Section 4.1 and Section 3.2"}],"minor_comments":[{"comment":"The abstract calls the multi-source experiments 'preliminary', but the introduction and conclusion claim scalability for multi-source localization as a key contribution. The paper should clarify whether multi-source localization is a demonstrated result or a preliminary observation.","section":"Abstract and Introduction"},{"comment":"The superscripts in r^{PHAT}_{ij} are rendered incorrectly in the text (appearing as 'rPHA T'), which makes the equation hard to read. This is likely a LaTeX formatting issue that should be fixed.","section":"Equation (2)"},{"comment":"The notation for loss weights is inconsistent: the text uses λsound, λm-loc, and λs-loc, while Eq. (8) uses λ_sound, λ_m-loc, and λ_s-loc. The authors should unify the notation.","section":"Section 3.2.3, Eq. (8)"},{"comment":"The entry for Multilat* at M=5 is listed as N/A, but no explanation is given. If the method cannot be run in that configuration, the reason should be stated.","section":"Table 2"},{"comment":"The multi-source localization results (MAE 45.21 cm for two sources and 69.86 cm for three sources) are reported in a single sentence with no comparison to baselines, no variance over runs, and no information about how the number of sources K is determined during inference. This is insufficient to support the scalability claim.","section":"Section 4.2"},{"comment":"Figure 2 is dense and the labels for the 'four highlighted components' mentioned in the caption are not clearly identifiable. The figure should be simplified or annotated more explicitly to help the reader map the components to the text.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The central issue is that Eq. (8) does not provide any source-localization supervision for the single-source case, yet the paper's main tables are single-source. This could be a typographical error, but without code or a corrected loss statement, the empirical results are unverifiable. The fault-tolerance claim is also under-supported. The paper appears to be a conference-style submission; if this is intended for a journal, the authors will need to substantially expand the method details, provide code, and add direct evaluations of source localization under faults. The current manuscript is not suitable for acceptance in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main thing to know: the stress-test note is correct, and it is the load-bearing issue. Eq. (8) defines the source-localization loss as a sum over source pairs (i,j)∈K. For K=1, that sum is empty, so the single-source results in Tables 1 and 2 have no stated training signal for the predicted source position. The reconstruction and microphone-coordinate losses do not supervise the source output. So the reported MAE and accuracy cannot be produced by the objective as written. This is an internal inconsistency in the central claim, not a philosophical disagreement.\n\nThat said, there is something here. The architecture makes sensible engineering choices: frozen BEATs embeddings, NGCC-PHAT TDOA features with adaptive coherence weighting, sparse cross-attention to cut computation, and a joint decoder for sources and faulty microphones. On the public LuViRA benchmark, the reported numbers beat wav2pos by 0.3–2.3 cm MAE, and the reduced-microphone table shows consistent gains. The microphone self-localization results are plausible and useful. If the loss were corrected and code released, this could be a solid applied contribution.\n\nThe soft spots beyond the equation are real but minor by comparison. The fault-tolerance claim is supported only by microphone-position MAE, not by source-localization error under faulty microphones. The paper does not report full hyperparameters, so independent reproduction is not currently possible. The multi-source results are preliminary and have no baselines, but that is secondary.\n\nWho is this for? Groups working on learning-based SSL with distributed mic arrays. They would find the architecture ideas worth building on, but they should not trust the absolute numbers until the training objective is clarified and code appears.\n\nRecommendation: send to peer review, but with a clear expectation of major revision. The loss equation must be fixed or the code released to show what was actually trained; fault-tolerant SSL should be evaluated by SSL error under faults. The paper deserves a serious referee because the benchmark comparisons are concrete and the architecture is not just noise, but it is not acceptable in its current form.","headline":"The architecture is promising and the benchmark numbers are plausible, but the stated training loss has no source-coordinate term for the single-source case, so the main results are not supported by the paper as written.","tokens_in":9038,"tokens_out":3418,"would_cite":false,"duration_ms":34158,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims a sparse cross-attention 3D sound localizer that uses frozen audio embeddings and weighted TDOA features beats prior methods on LuViRA music3/speech3 while needing fewer microphones and estimating faulty mic positions.","keywords":["sound source localization","3D localization","sparse cross-attention","microphone fault tolerance","masked autoencoder","TDOA features","adaptive signal coherence"],"falsifier":"Look at the training code for the K=1 case: if Eq. (8)'s source-localization term is the only localization loss, its gradient with respect to the predicted source position is zero, so the reported single-source MAE cannot be charged to that term. A direct experiment is to train the model on speech3 with and without an added absolute-coordinate term such as $|\\hat{r}-r|$ and compare the single-source MAE curves.","tokens_in":8076,"feed_emoji":"🎤","tokens_out":9566,"duration_ms":85632,"temperature":0.7,"pith_summary":"The paper tries to show that an end-to-end, learning-based 3D sound source localization system can be accurate, cheap, and robust at the same time. It proposes a model that fuses pretrained audio embeddings, microphone coordinates, and weighted time-difference-of-arrival features through sparse cross-attention, and trains it with masked reconstruction plus auxiliary losses. The claimed payoff is that the model beats existing localization methods on the LuViRA music3 and speech3 recordings while using fewer microphones, keeps working when some microphone positions are unknown or wrong, and can even estimate those faulty positions. A sympathetic reader would care because real deployments cannot always guarantee calibrated, dense, fully working arrays.","feed_headline":"Sparse-attention model beats 3D sound locators with fewer mics","feed_subtitle":"On LuViRA speech3, seven microphones match nine-microphone wav2pos accuracy, and faulty mic positions are estimated too.","key_machinery":"The object that carries the argument is a three-stream transformer-style network. The acoustic stream passes each microphone's raw audio through a frozen pretrained encoder (BEATs) to get audio embeddings; the coordinate stream encodes microphone positions with an MLP and extracts TDOA features via NGCC-PHAT, weighted by an adaptive signal coherence metric (ASCM) that down-weights unreliable pairs; the joint stream uses cross-attention to align audio with coordinates, then sparse cross-attention (retaining only the top-T attention weights per query) to fuse TDOA embeddings, a masked autoencoder to reconstruct masked audio and coordinate embeddings, and a position decoder to output source locations. The sparse attention is what keeps the model cheap as the number of microphones grows.","core_discovery":"On its own terms, the central discovery is that microphone-array localization does not need every microphone to be calibrated and reliable: the model jointly learns a spatial geometry from audio and coordinates, and in the paper's experiments it reports the best single-source MAE on LuViRA music3/speech3 (13.9 cm and 21.3 cm, with 96.8% and 88.4% within 30 cm), cuts the required array size (28.7 cm MAE with seven microphones, close to wav2pos's 28.4 cm with nine), localizes faulty microphone positions (down to 29.5 cm MAE in scene B with nine known microphones), and gives preliminary multi-source results of 45.21 cm and 69.86 cm MAE for two and three overlapping sources using the same eleven microphones.","pith_inferences":["If the missing single-source loss is resolved in the code, the reported gains may turn out to come mostly from the frozen BEATs embeddings and masked-reconstruction pretraining rather than from explicit geometric supervision; an ablation that adds a plain absolute-coordinate loss would show this.","The same architecture can be read as a self-calibrating array: by treating every microphone position as latent, the model could estimate an entire unknown array geometry from audio alone, which could be tested on ad-hoc recordings with moving microphones.","Because sparse cross-attention keeps only the top-T weights per query, the method should scale to much larger microphone counts than tested; a stress test with 20-50 microphones would show whether accuracy saturates or keeps improving.","The adaptive coherence weighting is a plausible source of robustness in reverberant rooms; a synthetic room-impulse-response sweep with varying reverberation time would separate its contribution from that of the pretrained encoder."],"forward_implications":["With seven microphones the model reaches 28.7 cm MAE on speech3, close to the 28.4 cm that wav2pos needs nine microphones to achieve, so deployment cost can drop.","Faulty microphones no longer force catastrophic failure: the model estimates an unknown microphone position to about 29.5 cm MAE in scene B with nine known microphones.","Preliminary multi-source results of 45.21 cm and 69.86 cm MAE for two and three simultaneous sources suggest the same hardware can handle overlapping sound events.","Inference takes 0.4 seconds with a 6.8M-parameter model, and the frozen audio encoder removes the need to train the acoustic front end."],"supporting_citations":[{"why":"Supplies the LuViRA music3/speech3 recordings and the training/validation split on which all main performance tables report.","marker":"[1]"},{"why":"Defines the wav2pos masked-autoencoder baseline that the proposed model outperforms, including the nine-microphone accuracy used as the comparison point.","marker":"[5]"},{"why":"Provides the shift-equivariant NGCC-PHAT features that form the TDOA input to the coordinate stream.","marker":"[6]"},{"why":"Supplies the frozen pretrained audio encoder whose embeddings are the acoustic input to the model.","marker":"[7]"},{"why":"Contributes the graph-neural-network baseline for distributed microphone arrays with variable microphone counts.","marker":"[4]"},{"why":"Provides the dual-input neural-network baseline for positional sound source localization.","marker":"[3]"},{"why":"Provides the robust multilateration baseline and TDOA self-calibration methods that motivate the calibration-free and fault-tolerant design.","marker":"[2]"}],"fun_headline_variants":["Sound locator thrives with fewer mics, even faulty ones","3D sound localization now handles mic faults","Sparse-attention SSL: efficient, fault-tolerant 3D","Multitask 3D sound mapping with broken mics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported single-source results depend on a training signal the paper does not actually supply: its source-localization loss (Eq. 8) sums over pairs of sources, so with a single source the sum is empty, and no alternative absolute-coordinate loss is stated.","fun_headline_variants_meta":{"raw":{"variants":["Sound locator thrives with fewer mics, even faulty ones","3D sound localization now handles mic faults","Sparse-attention SSL: efficient, fault-tolerant 3D","Multitask 3D sound mapping with broken mics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00029,"raw_usage":{"total_tokens":1647,"prompt_tokens":848,"completion_tokens":799,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":740}},"tokens_in":464,"tokens_out":799,"duration_ms":8132,"temperature":1.0,"reasoning_tokens":740,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:43:01.702326+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Look at the training code for the K=1 case: if Eq. (8)'s source-localization term is the only localization loss, its gradient with respect to the predicted source position is zero, so the reported single-source MAE cannot be charged to that term. A direct experiment is to train the model on speech3 with and without an added absolute-coordinate term such as $|\\hat{r}-r|$ and compare the single-source MAE curves.","supporting_citations":[{"cited_title":"Efficient and Microphone-Fault-Tolerant 3D Sound Source Localization","cited_arxiv_id":"2505.20961","evidence_quote":"Supplies the LuViRA music3/speech3 recordings and the training/validation split on which all main performance tables report."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the wav2pos masked-autoencoder baseline that the proposed model outperforms, including the nine-microphone accuracy used as the comparison point."},{"cited_title":"The luvira dataset: Synchronized vision, radio, and audio sensors for indoor localization,","cited_arxiv_id":null,"evidence_quote":"Provides the shift-equivariant NGCC-PHAT features that form the TDOA input to the coordinate stream."},{"cited_title":"Extension of time-difference-of-arrival self calibration solutions using ro- bust multilateration,","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen pretrained audio encoder whose embeddings are the acoustic input to the model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the graph-neural-network baseline for distributed microphone arrays with variable microphone counts."},{"cited_title":"Problem Definition Consider a 3D acoustic environment with M microphones at known coordinates rm ∈ R3, m = 1,","cited_arxiv_id":null,"evidence_quote":"Provides the dual-input neural-network baseline for positional sound source localization."},{"cited_title":"Traditional Sound Source Localization Traditional approaches to sound source localization are based on establishing geometric and statistical frameworks","cited_arxiv_id":null,"evidence_quote":"Provides the robust multilateration baseline and TDOA self-calibration methods that motivate the calibration-free and fault-tolerant design."}],"review_version":1}