{"id":"fba89f32-9c59-4a5b-9f7a-d56280029007","arxiv_id":"2412.13037","paper_version":7,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TAME applies parallel Mamba state-space models to audio spectrograms and reports state-of-the-art drone trajectory estimation and classification on MMAUD, with unresolved evaluation concerns.","lead":"Researchers built a drone-detection model that uses only sound to estimate a drone's 3D position and type. It reports top accuracy on the MMAUD benchmark, but the evaluation has gaps that make the headline claim only conditional.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central SOTA claim is unverifiable as described: §III-B sets J=16, W=4, H=1 for a 224×16 mel-spectrogram, but temporal patch splitting yields 56 temporal patches, not 16; and the audio-only day/night columns in Table I are identical, so the 'especially at night' claim is not supported.","rationale":"The reader's CONDITIONAL verdict remains appropriate. The single most load-bearing issue is not that audio-only SOTA is impossible in principle—Mamba on four-channel audio can plausibly exploit inter-channel delay—but that the described pipeline cannot be connected to the reported Table I numbers. Section III.A states 224×16 mel-spectrograms; Section III.B fixes W=4, H=1, J=16. With non-overlapping temporal splitting along 224 time frames, J_t=224/4=56, while the spectral sequence has J_s=16. Equations (1)-(2) use the single symbol J for both, so either the input size, the patch size, or the sequence length in the implementation differs from the manuscript. This does not prove the SOTA claim false; a corrected configuration could reproduce the table. It does make the central evidence internally inconsistent. The identical Light/Dark entries for all audio-only rows further mean no separate night evaluation is shown, so the qualitative claim 'especially at night' lacks experimental support. The open-source code is a credible route to settle this: if the forward pass with 224×16 and J=16, W=4, H=1 runs, or if a clearly documented alternative configuration reproduces Table I, the concern is resolved. I agree with the reader's conditional verdict but focus on the input-shape inconsistency as the load-bearing technical flaw rather than the day/night equivalence alone.","tokens_in":7360,"tokens_out":5768,"duration_ms":51926,"concrete_test":"Run the released GitHub code on MMAUD V1 with the prescribed 224×16 mel-spectrogram and J=16, W=4, H=1. First check whether the forward pass executes at all; if it requires a different input size or patch count, the published configuration is not the evaluated one. If it executes, recompute TAME's overline APE and Acc in Table I and compare them to 0.55 and 98.0. Separately, re-evaluate TAME on day-only and night-only audio subsets to determine whether the identical columns in Table I are actual measurements or placeholders.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's 'new standard' claim rests on Table I, where TAME reports APE 0.55 and Acc 98.0 against AV-FDTI at 1.24 and 94.0. For that comparison to be meaningful, the rows must describe the evaluated model. The manuscript's settings contradict this. §III.A fixes the mel-spectrogram at 224×16; if R is the temporal width (224) and S the spectral height (16), then §III.B's non-overlapping temporal patch split with W=4, H=1 yields J_t = R/W = 56 temporal patches, each of shape κ×4×16. The spectral split with H=1 yields J_s = S/H = 16 patches. The paper instead sets both to J=16 and writes both sequences as R^{J×(κWS)} and R^{J×(κRH)} in §II-A, so Eqs. (1)-(2) cannot simultaneously hold for the stated resolution. Additionally, Table I lists identical Light and Dark values for every audio-only row including TAME, and no day/night audio split is described; the claim of superiority 'especially at night' is therefore not established. Combined with the unexplained 'brightness attenuation' applied to an audio-only pipeline, the experimental protocol is too ambiguous to verify the headline result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes TAME, an audio-only model combining a temporal-spectral Mamba (TSMamba) backbone with a temporal feature enhancement (TFE) cross-attention module, for 3D drone trajectory estimation and drone-type classification on the MMAUD dataset. The model ingests multichannel audio mel-spectrograms, splits them into temporal and spectral patch sequences, processes each with selective state-space models, fuses them, and feeds a learned token to two MLP heads. The paper reports state-of-the-art results on MMAUD V1, with an average APE of 0.55 and classification accuracy of 98.0%, outperforming visual, audio, and audio-visual fusion baselines, and includes an ablation study and a public code release.","tokens_in":7698,"tokens_out":3892,"duration_ms":39109,"significance":"If the reported results are reproducible, the paper is significant: it suggests that a purely audio-based model with a state-space backbone can match or exceed audio-visual fusion methods on a recent multi-modal drone benchmark, which would be practically relevant for low-cost and lighting-invariant drone detection. The architecture is clearly motivated and the open-source release is a concrete strength. The main caveat is that the evaluation as reported is not yet reproducible: the patch-splitting configuration is internally inconsistent, the experimental protocol omits several key details, and the headline comparison rests on single-run numbers without uncertainty estimates. These issues are fixable, but they currently prevent full verification of the central claim.","major_comments":[{"comment":"The patch-splitting configuration is internally inconsistent. With a 224×16 mel-spectrogram and the stated J=16, W=4, H=1, the temporal patch sequence has R/W = 224/4 = 56 patches, not 16, while the spectral patch sequence has S/H = 16/1 = 16 patches. Equations (1) and (2) therefore cannot both use the same J for the two patch sequences at the stated input resolution. Please clarify which J applies to each stream, or correct the patch dimensions; this is load-bearing because the architecture shape and the reported results depend on the actual splitting scheme.","section":"§II-A, §III-B"},{"comment":"The evaluation protocol does not support the strength of the stated conclusions. The table reports a single run per method with no error bars, multiple seeds, or statistical tests, so the headline improvement (TAME APE 0.55 vs. AV-FDTI 1.24) could be within run-to-run variation. In addition, the audio-only rows report identical Light and Dark values for TAME, AudioNet, and DroneChase; while this is plausible for a lighting-invariant audio model, the text's claim of superiority 'especially at night' is not substantiated by any night-specific audio result, and the paper does not state how the day/night split is defined for audio recordings. Please provide uncertainty estimates and clarify the day/night audio protocol.","section":"Table I, §III-C"},{"comment":"The term 'Brightness attenuation' is applied during training/testing for an audio-only model, but the augmentation is never defined for audio or mel-spectrograms. It is unclear whether this is a visual augmentation erroneously carried over from an image pipeline, or an audio-specific transformation, and no parameters are given. This ambiguity prevents reproduction of the reported experimental pipeline and should be resolved by either defining the audio-domain augmentation precisely or removing the reference.","section":"§III-B"},{"comment":"The comparison omits Mamba-based audio baselines such as Audio Mamba [35] and SSAMBA [36], which are cited in the introduction but not evaluated. Since the contribution is specifically a Mamba-based audio architecture, including at least one such baseline would substantially strengthen the claim that the proposed temporal-spectral design, rather than the SSM backbone alone, is responsible for the gains. I regard this as a completeness issue that should be addressed in revision.","section":"§II-A, §III-C"}],"minor_comments":[{"comment":"The mel-spectrogram extraction is not fully specified; please state the STFT window size, hop length, number of mel filters, and whether the 224×16 dimensions correspond to time×frequency or frequency×time, as the patch-splitting description depends on this orientation.","section":"§III-A"},{"comment":"The dataset name is written as 'MMUAD' in the abstract but as 'MMAUD' elsewhere, including reference [41]; please unify the spelling.","section":"Abstract"},{"comment":"The table header is difficult to parse; the column grouping under Light, Dark, and Overline should be made explicit, and the 'Params' column should be clearly associated with the whole row rather than the Modality column.","section":"Table I"},{"comment":"The state-space model equations are not given; a brief description of the selective SSM update and the role of the learnable token would improve reproducibility, especially since the implementation is said to be public.","section":"§II-A"},{"comment":"The conclusion mentions 'reliance on large datasets' as a limitation, but the paper does not discuss the size of the training set or how the 7:3 split was performed; please add this information to the experimental section.","section":"§IV"}],"recommendation":"major_revision","confidential_remarks":"The patch-dimension inconsistency and the undefined audio augmentation are the main technical blockers; both are likely fixable with clarification and additional experimental details. The paper's central direction is timely and the open-source code is a positive feature, but the current manuscript is not reproducible as written. No concerns about authorship or data provenance surfaced from the manuscript itself."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. TAME is a serious attempt at audio-only drone 3D tracking and classification using a parallel temporal-spectral Mamba backbone with residual cross-attention fusion, and it ships code and trained models. But the headline \"new standard\" claim is not verifiable as written: the patch geometry in Section III-B does not match Section II-A (a 224×16 mel-spectrogram with W=4, H=1 gives 56 temporal patches, not 16), and the audio-only Light and Dark columns in Table I are identical, so \"especially at night\" is unsupported.\n\nWhat is genuinely new: the dual-axis patch design plus the Temporal Feature Enhancement neck is a legitimate combination, not present in Audio Mamba or SSAMBA, which apply Mamba to 1D audio tokens rather than two separate spectrogram axes. The ablation in Table II is internally consistent and shows TFE helps: APE drops from 0.68 with SFE to 0.55 with TFE, and TMamba alone beats SMamba alone. The open GitHub repository is a real asset; a referee can actually run the code.\n\nSoft spots, in order. First, the patch inconsistency is not cosmetic. If the implementation really uses J=16 for the temporal axis, the described horizontal split with W=4 is not what runs; if it uses 56 patches, the notation and Eq. (1) are wrong. The authors need to state the actual patch counts and align the published settings with the released code. Second, there are no error bars, multiple seeds, or significance tests anywhere, and the APE gap over AV-FDTI (0.55 vs 1.24) is large but could still be run-to-run variance without evidence. Third, the fact that every audio-only row has identical Light and Dark values suggests the dataset's audio conditions are not separated at all; the \"especially at night\" sentence overclaims. Fourth, \"brightness attenuation\" applied to an audio model is unexplained and sounds like leftover from a visual pipeline. Fifth, the dataset and several baselines come from the same NTU group, and there is no Mamba-based audio baseline; that limits novelty but is not a fatal circularity.\n\nWho this is for: researchers working on audio UAV detection or on Mamba spectrogram encoders. They will get a useful architecture and an independent benchmark to test against. The empirical claim is plausible but unproven.\n\nRecommendation: yes, send to peer review, but conditional. The authors should report exact patch dimensions and sequence lengths, provide error bars over multiple seeds, clarify how day and night audio are separated, and explain the brightness attenuation. The code makes this checkable. If the numbers hold, it is a solid engineering contribution; if not, the referee report is short.","headline":"Audio-only drone tracker with a sensible Mamba fusion idea, but the headline SOTA claim is unverifiable as reported until the patch geometry, the day/night audio rows, and the missing error bars are sorted out.","tokens_in":8199,"tokens_out":5833,"would_cite":false,"duration_ms":52504,"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":"An audio-only model estimates drone trajectories and classifies drone types better than visual and fusion baselines.","keywords":["audio-based drone detection","state space models","trajectory estimation","drone classification","mel-spectrogram","cross-attention","anti-UAV","temporal feature enhancement"],"falsifier":"Run the released code on MMAUD V1 with the stated settings and compare the reproduced APE and accuracy with the paper's 0.55 and 98.0; also check whether the day and night audio test files differ. If the numbers do not reproduce, or the day and night audio are identical, the central claim fails.","tokens_in":7164,"feed_emoji":"🎧","tokens_out":8965,"duration_ms":81502,"temperature":0.7,"pith_summary":"This paper introduces TAME, an audio-only system that estimates a drone's 3D trajectory and classifies the drone type from four channels of sound. The method turns multichannel audio into mel-spectrograms and runs parallel Mamba-style state-space scans along the time and frequency axes, then fuses the two streams with residual cross-attention. On the MMAUD V1 benchmark, the paper reports a mean trajectory error of 0.55 m and 98.0 percent classification accuracy, better than the published visual, audio, and audio-visual fusion methods it compares against. If those numbers hold, cheap passive microphones could serve as a standalone drone-threat sensor, including in conditions where cameras fail.","feed_headline":"Sound-only model beats camera and fusion drone trackers","feed_subtitle":"On the MMAUD benchmark, TAME reaches 98 percent classification accuracy and best trajectory error using only four audio channels.","key_machinery":"The load-carrying object is TSMamba, a parallel pair of selective state-space models applied to a mel-spectrogram. TMamba scans the spectrogram from left to right along time with a learnable token, capturing temporal-difference-of-arrival features; SMamba scans top to bottom along frequency, capturing global spectral attenuation. The Temporal Feature Enhancement module then injects spectral information into temporal features via residual cross-attention, and the enhanced temporal token is read out by two MLP heads—one for 3D trajectory regression with L1 loss and one for drone-type classification with cross-entropy loss.","core_discovery":"The central claim is that a temporal–spectral Mamba backbone, called TSMamba, can extract from audio alone the information needed to locate a drone in 3D and identify its model. The paper's mechanism is that a temporal scan of the mel-spectrogram encodes time-difference-of-arrival features that carry distance and, in part, category information, while a spectral scan encodes the sound-attenuation profile across frequencies. A Temporal Feature Enhancement module merges the spectral features into the temporal features using cross-attention with a residual connection, so that a single learnable temporal token carries the information used by two lightweight heads. TAME is reported to reach 0.55 average position error and 98.0 percent accuracy on MMAUD V1, outperforming the strongest fusion baseline AV-FDTI at 1.24 average position error and 94.0 percent accuracy.","pith_inferences":["The paper does not report how performance varies with distance or noise level; a natural next experiment is to measure trajectory error in decibel bins to see where audio-only estimation starts to fail.","A stronger test of the paper's 'especially at night' framing would record identical drone flights in true day and night conditions, since the current audio rows do not vary with illumination.","The architecture could be transferred to nearby problems such as speaker localization or acoustic event direction-finding, where the temporal-difference and spectral-attenuation features have the same physical meaning."],"forward_implications":["Audio-only detection could make drone monitoring systems smaller, cheaper, and harder to detect, since they need only a microphone array rather than cameras or radar.","The method's night-time numbers being equal to its day-time numbers implies the audio features are treated as illumination-independent; if that holds, audio-only systems would remain useful exactly where visual trackers fail.","Because the model is a generic temporal/spectral sequence learner, the same TSMamba-TFE design can be reapplied to other tasks that need location and category from one sensor stream.","Releasing code and trained weights lets any group rerun the MMAUD V1 comparison, so the claimed lead over fusion baselines is directly checkable."],"supporting_citations":[{"why":"Supplies the MMAUD anti-UAV benchmark, the four-channel audio recordings, and the ground-truth 3D trajectories used in every experiment.","marker":"[41]"},{"why":"Defines the selective state-space model that TSMamba is built from and that gives the method its linear-time sequence processing.","marker":"[39]"},{"why":"AV-FDTI is the strongest audio-visual fusion baseline that TAME claims to beat on trajectory error and classification accuracy.","marker":"[29]"},{"why":"AV-PEDaware supplies both the audio baseline AudioNet and the visual baseline VisualNet whose published numbers appear in the comparison table.","marker":"[37]"},{"why":"DroneChase is the audio-only tracking baseline whose trajectory and classification numbers TAME is compared against.","marker":"[28]"},{"why":"TalkNet is the audio/active-speaker-detection baseline included in the comparison for trajectory estimation and classification.","marker":"[38]"},{"why":"Audio Mamba is the earlier state-space audio representation work that motivates applying Mamba to audio spectrograms.","marker":"[35]"},{"why":"ViT provides the learnable-token and patch-embedding design that TAME adapts for splitting spectrograms along time and frequency.","marker":"[40]"},{"why":"DarkNet is the visual-detection baseline whose day and night numbers are used as the visual comparison point.","marker":"[42]"}],"fun_headline_variants":["Audio-only Mamba nails drone 3D paths and types","Sound-only TAME beats fusion trackers in drone task","Mamba listens: audio-only drone tracking beats baselines","TSMamba: audio-only 3D trajectory and classification wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed results depend on the experiments being run exactly as described, including genuinely separate day and night audio recordings; if the description and the actual run diverge, the comparison against other methods is not yet established.","fun_headline_variants_meta":{"raw":{"variants":["Audio-only Mamba nails drone 3D paths and types","Sound-only TAME beats fusion trackers in drone task","Mamba listens: audio-only drone tracking beats baselines","TSMamba: audio-only 3D trajectory and classification wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000255,"raw_usage":{"total_tokens":1535,"prompt_tokens":870,"completion_tokens":665,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":595}},"tokens_in":486,"tokens_out":665,"duration_ms":6763,"temperature":1.0,"reasoning_tokens":595,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:28:44.420538+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code on MMAUD V1 with the stated settings and compare the reproduced APE and accuracy with the paper's 0.55 and 98.0; also check whether the day and night audio test files differ. If the numbers do not reproduce, or the day and night audio are identical, the central claim fails.","supporting_citations":[{"cited_title":"MMAUD: A Comprehensive Multi-Modal Anti-UA V Dataset for Modern Miniature Drone Threats,","cited_arxiv_id":null,"evidence_quote":"Supplies the MMAUD anti-UAV benchmark, the four-channel audio recordings, and the ground-truth 3D trajectories used in every experiment."},{"cited_title":"Av- fdti: Audio-visual fusion for drone threat identification,","cited_arxiv_id":null,"evidence_quote":"AV-FDTI is the strongest audio-visual fusion baseline that TAME claims to beat on trajectory error and classification accuracy."},{"cited_title":"Av-pedaware: Self-supervised audio-visual fusion for dynamic pedestrian awareness,","cited_arxiv_id":null,"evidence_quote":"AV-PEDaware supplies both the audio baseline AudioNet and the visual baseline VisualNet whose published numbers appear in the comparison table."},{"cited_title":"Dronechase: A mobile and automated cross-modality system for continuous drone tracking,","cited_arxiv_id":null,"evidence_quote":"DroneChase is the audio-only tracking baseline whose trajectory and classification numbers TAME is compared against."},{"cited_title":"Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection,","cited_arxiv_id":null,"evidence_quote":"TalkNet is the audio/active-speaker-detection baseline included in the comparison for trajectory estimation and classification."},{"cited_title":"An image is worth 16x16 words: Transformers for image recognition at scale,","cited_arxiv_id":null,"evidence_quote":"ViT provides the learnable-token and patch-embedding design that TAME adapts for splitting spectrograms along time and frequency."}],"review_version":1}