{"id":"4b735ca9-6828-4fff-8bf4-47978120c88d","arxiv_id":"2608.00875","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A drone-noise suppression method that lets separation masks exceed unity and adds a residual correction improves downstream classification of human vocal sounds in low-SNR drone recordings.","lead":"This paper presents a drone noise suppression model that estimates rotor noise with a source separation network and then subtracts it to recover human sounds. It shows its largest gains for human vocal sounds in very noisy drone recordings, pointing toward better drone-based search and rescue.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline 'statistically significant' HV gains rest on unadjusted pairwise t-tests over non-independent multi-channel segments; no clustering or multiple-comparison correction is reported.","rationale":"The reader's verdict is CONDITIONAL, and the reader's rationale already flags the absence of multiple-comparison adjustment and the overclaim of 'consistently improves' for non-vocal categories. My stress-test identifies the same cluster of issues but locates the load-bearing point more precisely: the headline statistical significance is the main support for the central claim, and that support depends on an unreported and probably invalid inference procedure. The OOD generalization concern raised by the reader is legitimate but is empirically supported by Table 3 and Figure 3, so I do not rank it as the single most load-bearing issue. The ablation also shows that the additive complex residual contributes almost nothing beyond the alpha scaling (0.73 vs 0.73 at -20 to -10 dB; 0.17 vs 0.16 at -30 to -20 dB), which further weakens the claim that both proposed mechanisms matter, though it does not destroy the overall system-level result. These concerns do not justify rejection: the paper provides a useful benchmark and a plausible empirical direction, and the requested statistical re-analysis and claim-narrowing are achievable in revision. Hence the verdict remains CONDITIONAL, i.e., unchanged from the reader's verdict.","tokens_in":11472,"tokens_out":7940,"duration_ms":77336,"concrete_test":"Obtain the per-segment F1 predictions and the mapping from each segment to its source recording, microphone channel, and DroneAudioSet configuration. Recompute the AudioSep-FT versus DRONEAUDIONET comparisons as cluster-robust paired tests, clustering by source recording or configuration, and apply Benjamini-Hochberg correction at q=0.05 across all nine category-by-SNR F1 tests in Table 1 plus the three OOD tests in Table 3. Report adjusted p-values and 95% confidence intervals for the HV differences; if the -20 to -10 dB and -30 to -20 dB gains remain significant under this procedure, the statistical objection is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central positive claim is the statistically significant improvement for human vocal sounds in Table 1 (HV: 0.73 vs 0.66 at -20 to -10 dB; 0.17 vs 0.13 at -30 to -20 dB), marked with dagger markers from pairwise t-tests between AudioSep-FT and DRONEAUDIONET. The testing procedure is not described in sufficient detail to support this claim. DroneAudioSet records each configuration simultaneously across multiple microphones, and Section 4.1 says each channel is treated independently 'resulting in increased spatial diversity.' If the unit of analysis is a 5-second segment, then the same physical source event appears in multiple channels and possibly in overlapping segments, so the observations are not independent. A plain t-test over such segments overstates the effective sample size and can produce spuriously small p-values. No sample size, pairing scheme, clustering variable, or multiple-comparison adjustment is reported, despite at least nine F1 comparisons in Table 1, three in Table 3, and many mode/class comparisons in Figure 3. Under this testing protocol, selecting two significant HV bands from many comparisons is not convincing evidence. Additionally, Table 1 itself contradicts the abstract's 'consistently improves': DRONEAUDIONET is significantly worse than AudioSep-FT for Non-Human sounds at -30 to -20 dB (0.21 vs 0.24) and numerically worse in several HNV/NH rows. The claim should therefore be narrowed to category- and SNR-specific gains, and those gains need correction for clustering and multiplicity before they can carry the headline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"DRONEAUDIONET adapts the query-based separation model AudioSep to drone noise suppression by estimating the drone component with the fixed text query 'Drone motor and propellor sounds' and subtracting it from the mixture. The two proposed architectural changes are a learnable scalar α that multiplies the sigmoid-bounded mask, allowing mask magnitudes above one, and an additive complex residual β|R|e^{j∠R} predicted by extra network output channels. The model is initialized from AudioSep and fine-tuned on DroneAudioSet; evaluation is on held-out DroneAudioSet natural mixtures and on a DREGON-based out-of-domain set, using SI-SDR and F1 of a frozen SSLAM classifier. The main reported finding is a statistically significant improvement in classification F1 for human vocal sounds at low SNR (0.66 to 0.73 at -20 to -10 dB; 0.13 to 0.17 at -30 to -20 dB), with smaller or negative effects for other categories. The paper also provides ablations, mask visualizations, and resource usage tables.","tokens_in":11748,"tokens_out":8266,"duration_ms":68982,"significance":"The paper is a solid applied contribution with reproducible elements: it uses public datasets, a held-out natural test set, an OOD test, three public baselines plus a fine-tuned baseline, and an ablation that isolates the mask-scaling mechanism. The cIRM analysis (Figure 2) provides principled motivation for allowing mask magnitudes above unity in drone-dominant mixtures, and the released code and website support reproducibility. However, the headline claim of consistent improvement is not supported by the full result table, and the statistical evidence for the main gains is currently under-specified. If the statistical analysis is corrected and the claims are scoped to the HV category, the paper would be a useful benchmark in drone audition; as it stands, the reported gains over the fine-tuned AudioSep baseline are modest and category-specific.","major_comments":[{"comment":"The significance markers in Tables 1 and 3 and Figure 3 are based on unadjusted pairwise t-tests that appear to treat each 5-second segment of each microphone channel as an independent observation. DroneAudioSet records each configuration across multiple microphones simultaneously (Section 4.1), so segments from the same physical event and the same drone are correlated; treating them as independent inflates the effective sample size and can produce spurious p-values. The manuscript reports no sample sizes, no pairing information, no clustering variable, and no multiple-comparison correction, despite at least nine F1 comparisons in Table 1, three in Table 3, and many mode/class comparisons in Figure 3. Please replace these tests with cluster-robust or mixed-effects analyses (e.g., clustering by recording configuration or by drone platform and source event), or at minimum report the number of independent recordings and apply a multiple-comparison correction. Without this, the central claim of statistically significant improvement for HV is not established.","section":"§4.1, Table 1 footnote, §5.3"},{"comment":"The abstract's claim that DRONEAUDIONET 'consistently improves downstream sound classification performance' is contradicted by Table 1(b), where DRONEAUDIONET is numerically worse than AudioSep-FT for HNV at -10 to 0 dB (0.69 vs 0.72) and for NH in all three bands (-10 to 0: 0.40 vs 0.44; -20 to -10: 0.38 vs 0.40; -30 to -20: 0.21 vs 0.24, the last marked as a significant advantage for AudioSep-FT). The narrative should be narrowed to category- and SNR-specific gains, mainly for human vocal sounds in the lower two SNR bands, and the significant NH degradation should be acknowledged.","section":"Abstract and Table 1(b)"},{"comment":"In Table 1(a), DRONEAUDIONET is numerically worse than AudioSep-FT in SI-SDR for the HV category in all three SNR bands (e.g., -9.71 vs -9.54 dB at -20 to -10 dB; -21.68 vs -21.64 dB at -30 to -20 dB). Section 5.1's statement that 'SI-SDR improvements over AudioSep-FT remain modest and are generally comparable' is therefore misleading for the headline category: the proposed model does not improve reconstruction fidelity over the fine-tuned baseline. The paper should state explicitly that the benefit is in downstream classification, not in waveform-level SI-SDR.","section":"§5.1, Table 1(a)"},{"comment":"Table 2 shows that the full complex residual branch adds essentially nothing over the learnable α by itself: at -20 to -10 dB the F1 is 0.73 for both the α-only variant and DRONEAUDIONET, and at < -20 dB it is 0.16 vs 0.17. The only clear benefit of the residual branch appears in the < -20 dB band, and it is within one hundredth. This weakens the paper's claim that the additive complex residual is a complementary corrective mechanism (Section 3.2); the authors should either provide a stronger analysis (e.g., significance tests for the ablation differences, or a version without α) or temper the claim about the residual's contribution.","section":"§5.2, Table 2"}],"minor_comments":[{"comment":"There are several typos and formatting artifacts: 'UA Vs' appears in the Abstract and Introduction, 'propellor' is used instead of 'propeller', and 'V ocal' appears with a stray space in multiple places. Please proofread the camera-ready version.","section":"Abstract, Introduction"},{"comment":"The retention threshold 'input SNR ≥ −30 dB' is not justified. Since the paper targets extreme low-SNR conditions, please clarify why configurations below -30 dB were excluded and whether the main conclusions are sensitive to this choice.","section":"§4.1"},{"comment":"Fine-tuning details are incomplete: the manuscript does not report the learning rate, batch size, optimizer, number of epochs actually selected, or the validation-based early stopping criterion used for DRONEAUDIONET and AudioSep-FT. These details are needed for reproducibility.","section":"§4.3"},{"comment":"The caption states 'Dregon OOD' rather than 'DREGON OOD', and the text in §5.3 refers to 'statistically significant gains in several cases' without identifying which modes and classes reached significance; please list these explicitly in the caption or main text.","section":"Figure 3"},{"comment":"The SI-SDR evaluation takes the maximum over shifts of up to ±1000 samples (62.5 ms). This is an unusually large alignment window for SI-SDR and may mask reconstruction errors; please justify the 62.5 ms allowance and report results for the zero-shift case as well.","section":"§4.2"},{"comment":"The footnote assigns the dagger to the 'better-performing model,' which means a dagger can indicate a significant advantage for either method. Because the paper's focus is DRONEAUDIONET, it would be clearer to use two marker styles (e.g., upward and downward triangles) or to report the p-values for all comparisons in a supplementary table.","section":"Table 1 footnote"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope. The core idea (unbounded mask scaling plus a complex residual for known-noise estimation) is plausible and the empirical setup is reasonable, but the current significance analysis and the overbroad 'consistently improves' claim are not acceptable as-is. I would not reject, because the issues are fixable within the manuscript's scope: cluster-robust or mixed-effects testing, a multiple-comparison correction, and a carefully scoped abstract/introduction. I also note that the OOD set is only out-of-domain in drone hardware and flight mode, not in the source sounds, so the introduction's 'generalizes effectively' wording should be scoped accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is the short version: the paper is worth engaging with as a benchmark and as a demonstration that reframing AudioSep as a drone-noise estimator can help recover human vocal sounds in very low SNR. But the central claim is overstated, and the statistical evidence for the headline improvements is weaker than the prose suggests.\n\nWhat is genuinely good: the cIRM analysis showing that nearly half of drone-mask values exceed unity is a clean empirical motivation for unbounded masks. The two architectural modifications (learnable scalar alpha and additive complex residual) are simple, clearly explained, and ablated. The DREGON out-of-domain test is a real strength, and the careful use of public datasets and reporting of compute/resources helps reproducibility. The paper also includes a sensible limitation statement acknowledging hovering-only training and single-channel operation.\n\nThe soft spots matter. The abstract says the method 'consistently improves downstream sound classification performance,' but Table 1(b) contradicts that for HNV and NH: AudioSep-FT is better in several rows, sometimes by a wide margin (HNV at -10 to 0: 0.72 vs 0.69; NH at -10 to 0: 0.44 vs 0.40). The SI-SDR numbers for HV are also slightly worse for DRONEAUDIONET in every SNR band. So the claim should be narrowed to category- and SNR-specific gains for human vocal sounds, not 'consistently improves.'\n\nThe statistical support for the significant HV gains is the biggest concern. The daggers come from pairwise t-tests between AudioSep-FT and DRONEAUDIONET, but the paper treats every 5-second segment and every microphone channel as independent. DroneAudioSet records each configuration across multiple mics simultaneously, so segments from the same physical event are correlated. The paper reports no clustering variable, no sample size, no pairing scheme, and no multiple-comparison adjustment, despite at least nine comparisons in Table 1, three in Table 3, and many in Figure 3. Plain t-tests over dependent observations will inflate significance. The effect sizes are also modest (0.73 vs 0.66; 0.17 vs 0.13), so the real question is whether the direction survives a proper clustered analysis with correction for multiplicity.\n\nI also want more detail on the exact data splits and training hyperparameters; the paper mentions train/validation/test hours but not the number of configurations per split, and it does not say whether any recordings from the same drone platform appear in both training and test. That matters for the OOD claim.\n\nBottom line: this is a legitimate empirical contribution that should go to peer review, but with major revision. The abstract needs to be reined in, the statistical analysis needs to account for dependence and multiplicity, and the authors need to show whether the HV gains still hold. If they do, the paper will be a solid benchmark for drone audition. If not, the value of the architectural tweaks alone is modest. I would engage with it as a referee and I would probably cite it for the benchmark even now.","headline":"A useful drone-audition benchmark and plausible noise-estimation tweaks, but the headline gains rest on shaky statistics and an overbroad abstract claim.","tokens_in":12322,"tokens_out":1827,"would_cite":true,"duration_ms":18496,"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":"DRONEAUDIONET claims that reframing AudioSep as a drone-noise estimator with an unbounded mask and an additive complex residual consistently improves downstream sound classification in low-SNR drone recordings, most for human vocal sounds.","keywords":["drone noise suppression","source separation","audio query conditioning","drone audition","search and rescue","low SNR","mask scaling","residual correction"],"falsifier":"Record a human speaker under a drone model not present in DroneAudioSet or DREGON (for example, a fixed-wing or coaxial rotorcraft), mix at -20 dB SNR, and compare DRONEAUDIONET's downstream F1 for the recovered speech against AudioSep-FT; if it does not beat the fine-tuned baseline, the claim that the unbounded-mask and residual mechanism generalizes across unseen drone hardware fails.","tokens_in":1728,"feed_emoji":"🚁","tokens_out":1740,"duration_ms":58649,"temperature":0.7,"pith_summary":"The paper tries to show that the right way to recover sounds recorded from a drone is to estimate the drone's own rotor noise and subtract it, rather than trying to isolate the target sound directly. It adapts a language-query source-separation model to act as a drone-noise estimator, letting the estimated mask exceed magnitude one and adding a learned complex residual. Across DroneAudioSet and out-of-domain DREGON mixtures, the method consistently improves downstream sound classification, with the largest gains for human speech and cries at SNRs between -30 and -10 dB. If true, it gives search-and-rescue teams a more practical way to hear vulnerable people under severe rotor interference.","feed_headline":"Drone-noise suppressor lifts voice detection at low SNR","feed_subtitle":"Unbounded masks plus residual correction recover human vocal sounds better than fine-tuned AudioSep on drone recordings.","key_machinery":"The central object is the complex ideal ratio mask $D/X$ for the drone component in a mixture $X$. The paper shows that in drone-dominant mixtures, nearly half of these values exceed the unit circle, so a sigmoid-bounded mask systematically underestimates the drone. DRONEAUDIONET carries the argument with two modifications: a learnable scalar $\\alpha$ applied to the magnitude mask, which relaxes the bound, and an additive complex residual $\\beta|R|e^{j\\angle R}$, which corrects what the mask misses. The recovered source is the mixture minus the estimated drone signal, $\\hat{s} = x - \\hat{d}$, and the fixed text query 'Drone motor and propellor sounds' is embedded by CLAP and conditions the ResUNet to estimate the drone signal.","core_discovery":"On its own terms, the discovery is that the dominant obstacle in drone audition is not the unknown target but the known drone, and that standard mask-based separators fail because their bounded sigmoid mask cannot represent drone-dominated mixtures. In DroneAudioSet test mixtures, 49.3% of the ideal complex ratio mask values for the drone component lie outside the unit circle, whereas only 0.25% do for the source. DRONEAUDIONET therefore replaces the bounded mask with a learnable scalar $\\alpha$ that multiplies the sigmoid mask, and adds a residual term $\\beta|R|e^{j\\angle R}$ with learned complex values. Fine-tuned from AudioSep, it raises downstream classification F1 for human vocal sounds from 0.66 to 0.73 at -20 to -10 dB SNR and from 0.13 to 0.17 at -30 to -20 dB SNR, with similar gains out-of-domain on DREGON (0.69 versus 0.63).","pith_inferences":["Editorial inference: if the language-query conditioning is what lets the estimator generalize, replacing the fixed drone query with richer multi-word queries or acoustic reference embeddings could extend the approach to other structured noise sources such as wind or rain.","Editorial inference: the learned $\\alpha$ stays near 1.02 and $\\beta$ below 1, so one could test whether making these scalars frequency-dependent or per-band yields further low-SNR gains without destabilizing training.","Editorial inference: the paper's own future-work direction suggests that jointly fine-tuning the enhancer with the downstream classifier could push F1 higher in the hardest -30 to -20 dB regime.","Editorial inference: a natural stress test is to evaluate at SNRs below -30 dB and on transient events such as footsteps and clapping, where the paper reports little or no improvement."],"forward_implications":["At -20 to -10 dB SNR, downstream classification F1 for human vocal sounds rises from 0.66 to 0.73, meaning more speech and cries become detectable in realistic search-and-rescue conditions.","The gains extend to out-of-domain recordings from DREGON with unseen drone hardware and flight modes, especially for human vocal and non-human sounds.","Ablations show that the learnable mask scale $\\alpha$ is the main driver of improvement, while the complex residual adds a smaller corrective effect; removing residual phase hurts performance.","Because the method recovers the source by subtraction rather than by matching a target class, it remains class-agnostic and can preserve open-domain sounds such as alarms, cries, and mechanical events.","Waveform-level reconstruction gains are modest, but the method's benefit shows up in semantic downstream recognition, suggesting that signal fidelity metrics alone understate its practical value."],"supporting_citations":[{"why":"The base source-separation model that DRONEAUDIONET reframes as a drone-noise estimator and fine-tunes from AudioSet pretraining.","marker":"[11]"},{"why":"The DroneAudioSet dataset supplies aligned drone-only, source-only, and drone-with-source recordings used for training and in-domain evaluation.","marker":"[7]"},{"why":"The DREGON dataset provides the out-of-domain test set with unseen drone hardware and flight modes.","marker":"[21]"},{"why":"The decoupling of magnitude and phase estimation with a ResUNet motivates the additive complex residual branch.","marker":"[8]"},{"why":"The complex ideal ratio mask formulation is the basis for the unit-circle analysis motivating unbounded mask magnitudes.","marker":"[24]"},{"why":"SSLAM serves as the frozen self-supervised classifier that turns recovered audio into the downstream F1 metric.","marker":"[1]"},{"why":"The CLAP encoder produces the embedding for the fixed text query that conditions drone-noise estimation.","marker":"[25]"}],"fun_headline_variants":["Unbounded masks beat drone roar for voice rescue","Drone noise suppressor boosts voice detection at -20 dB SNR","Mask scaling plus residual fixes drone-dominated audio","DRONEAUDIONET: turning source separation into drone noise estimation","Unbounded mask trick lifts drone-audio voice recognition"],"cache_read_input_tokens":14336,"weakest_assumption_plain":"The method assumes the fixed text query 'Drone motor and propellor sounds' produces a CLAP embedding that captures the acoustic signature of any drone the system encounters; if that embedding misses an unseen drone's rotor noise, the estimated drone signal is wrong and subtraction will not recover the source.","fun_headline_variants_meta":{"raw":{"variants":["Unbounded masks beat drone roar for voice rescue","Drone noise suppressor boosts voice detection at -20 dB SNR","Mask scaling plus residual fixes drone-dominated audio","DRONEAUDIONET: turning source separation into drone noise estimation","Unbounded mask trick lifts drone-audio voice recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00078,"raw_usage":{"total_tokens":3453,"prompt_tokens":957,"completion_tokens":2496,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2415}},"tokens_in":573,"tokens_out":2496,"duration_ms":16076,"temperature":1.0,"reasoning_tokens":2415,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:15:58.650608+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a human speaker under a drone model not present in DroneAudioSet or DREGON (for example, a fixed-wing or coaxial rotorcraft), mix at -20 dB SNR, and compare DRONEAUDIONET's downstream F1 for the recovered speech against AudioSep-FT; if it does not beat the fine-tuned baseline, the claim that the unbounded-mask and residual mechanism generalizes across unseen drone hardware fails.","supporting_citations":[{"cited_title":"SSLAM: Enhancing self-supervised models with audio mixtures for polyphonic soundscapes","cited_arxiv_id":null,"evidence_quote":"SSLAM serves as the frozen self-supervised classifier that turns recovered audio into the downstream F1 metric."},{"cited_title":"Separate anything you describe.IEEE Transactions on Audio, Speech and Language Processing, 33:458–471, 2024","cited_arxiv_id":null,"evidence_quote":"The base source-separation model that DRONEAUDIONET reframes as a drone-noise estimator and fine-tunes from AudioSet pretraining."},{"cited_title":"Droneaudioset: An audio dataset for drone-based search and rescue.Neurips, 2025","cited_arxiv_id":null,"evidence_quote":"The DroneAudioSet dataset supplies aligned drone-only, source-only, and drone-with-source recordings used for training and in-domain evaluation."},{"cited_title":"Dregon: Dataset and meth- ods for uav-embedded sound source localization","cited_arxiv_id":null,"evidence_quote":"The DREGON dataset provides the out-of-domain test set with unseen drone hardware and flight modes."},{"cited_title":"Complex ratio masking for monau- ral speech separation.IEEE/ACM transactions on audio, speech, and language processing, 24(3):483–492, 2015","cited_arxiv_id":null,"evidence_quote":"The complex ideal ratio mask formulation is the basis for the unit-circle analysis motivating unbounded mask magnitudes."},{"cited_title":"Large-scale contrastive language-audio pretraining with feature fusion and keyword- to-caption augmentation","cited_arxiv_id":null,"evidence_quote":"The CLAP encoder produces the embedding for the fixed text query that conditions drone-noise estimation."}],"review_version":2}