{"id":"ab5c32de-d1e1-440c-a20d-1cf8e0d126ef","arxiv_id":"2608.10054","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"A U-Net trained on ridge-synthesized pseudo-clean mouse USVs, with a ridge-weighted loss, improves SI-SDR, ridge tracking, and classification accuracy on noisy bioacoustic recordings.","lead":"This paper builds a denoiser for noisy animal recordings that needs no clean audio for training, using automatically tracked vocalization contours to synthesize training data and to guide the loss. In mouse ultrasonic recordings, the method improves signal quality, contour tracking, and downstream syllable classification over several baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SI-SDR and ridge-tracking gains are measured against targets/annotations derived from the same ridge tracker used to synthesize training data; the more independent classification evidence is modest and lacks error bars.","rationale":"The reader's weakest_assumption concerns the accuracy of the tracked ridges used as clean training targets. My concern is adjacent but more specific: the evaluation of the two headline metrics (SI-SDR and ridge-tracking accuracy) is partly circular because the same ridge-tracking and ridge-mask synthesis pipeline generates both the training targets and the references against which success is measured. The classification results, especially on the domesticated strain, are the strongest non-circular evidence, and they do support the central claim, but the gains are small and reported without variance. This justifies keeping the CONDITIONAL verdict rather than rejecting the paper. I recommend UNCHANGED because the reader already identified the core risk and assigned CONDITIONAL with high confidence; my stress-test sharpens the circularity argument and adds a concrete falsification test, but does not move the verdict. The unspecified b in Eq. (5) is a real reproducibility issue, but it is secondary to the evaluation-circularity concern and would be a minor revision rather than a reason to change the overall verdict.","tokens_in":24229,"tokens_out":5365,"duration_ms":60530,"concrete_test":"Build a synthetic test set whose clean signals are generated from independent ground-truth ridges, e.g., manually annotated from scratch by annotators blinded to the automatic tracker, rather than from the tracker's own output. Add the same real noise types used in Section IV-B2, run the proposed denoiser, and recompute SI-SDR against these independent references. In parallel, on a random subset of the 193 annotated USVs, have annotators label ridges from scratch without seeing the tracker output, and recompute Table II and Table III precision/recall/deviation. If the SI-SDR gains and ridge-tracking improvements persist under these independent references, the concern is resolved. If they shrink substantially, the headline should be re-framed to rely on the classification results only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claims rest partly on evaluation loops that share the ridge-extraction pipeline used to build the training targets. The SI-SDR test set (Section IV-C) is synthesized from the same ridge-mask inversion procedure as the training targets (Sections III-C and III-D), so a U-Net trained to reproduce that generator can achieve large SI-SDR improvements without recovering the true clean signal. The paper acknowledges this as a 'controlled, relative comparison', but the conclusion still reports SI-SDR as headline evidence. The ridge-tracking evaluation is similarly entangled: the manual annotations were 'initially estimated with [16] and subsequently manually refined post hoc' (Section IV-C), and [16] is also the tracker applied to the denoised outputs and the tracker used in training-set synthesis. Thus, improved ridge-tracking scores can partly reflect the denoiser conforming to the tracker's own bias rather than to independent ground truth. The least entangled evidence is the classification improvement on out-of-sample domesticated data (macro-F1 57 to 61) and incorrect-ridge wild data (69 to 72), but these gains are small, no error bars or seed variance are reported, and the denoiser and classifier are trained on the same wild-mouse corpus, so the out-of-sample wild test is not fully independent. If the circular SI-SDR and ridge numbers are set aside, the remaining independent evidence is a 3-4 point F1 gain without uncertainty quantification. A separate reproducibility gap is that Eq. (5) never specifies b, the half-width of the ridge-loss weighting band, which is needed to reproduce the central loss.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a supervised U-Net denoiser for bioacoustic recordings in settings where no clean training data are available. The authors synthesize clean training targets from noisy mouse USV recordings by tracking fundamental and first-harmonic ridges with a multicomponent ridge-tracking method, generating a Gaussian mask around the ridges, and inverting the masked reassigned spectrogram. The same ridge information is used to define a ridge-guided loss that weights the training objective toward vocalization regions. The denoiser predicts a complex ratio mask and is evaluated in three ways: SI-SDR on a synthetic test set, fundamental and harmonic ridge-tracking accuracy on real field recordings with manually refined annotations, and USV classification macro-F1 using a BootSnap classifier retrained on denoised data from wild and domesticated mice. The paper reports large SI-SDR gains (e.g., from −10.17 dB to 16.52 dB at the lowest input SNR), reduced fundamental ridge-tracking frequency deviation (from 1.4 ± 3.4 kHz to 0.8 ± 1.5 kHz), and improved macro-F1 scores (e.g., from 83% to 89% on in-sample wild data and from 57% to 61% on out-of-sample domesticated data).","tokens_in":24328,"tokens_out":5116,"duration_ms":47541,"significance":"If the reported results hold, the paper addresses a genuine practical problem in bioacoustics: training deep denoisers without clean references. The idea of using automatically tracked ridges both for target synthesis and for loss weighting is novel and, because the method is not tied to one ridge tracker, it is potentially transferable to other tonal bioacoustic signals. The authors also provide publicly available code, a GUI, and a manually annotated ridge dataset, which is a concrete contribution. However, the evaluation is substantially entangled with the ridge-tracking pipeline used to build the training targets, and the most independent evidence (classification) is modest and lacks uncertainty quantification. The significance of the central claim therefore depends on whether the authors can add evaluations that break the self-referential loop, especially for SI-SDR and ridge tracking.","major_comments":[{"comment":"The synthetic test set for SI-SDR is generated using the same ridge-tracking and masked-reassignment inversion procedure (Sections III-C and III-D) that is used to synthesize the training targets. Consequently, a U-Net trained to approximate that target generator can achieve large SI-SDR gains on this test set without recovering the true clean signal; the 'clean' reference is itself an output of the training-target pipeline. The authors acknowledge this when they write that 'SI-SDR results should be interpreted as a controlled, relative comparison under synthetic conditions,' but the abstract and conclusion still cite the SI-SDR numbers as headline evidence of denoising quality. To support the central claim, the authors should either include an independent synthetic or semi-synthetic test set (e.g., clean signals synthesized from independently validated ridges, or real clean recordings with added noise) or explicitly remove SI-SDR from the list of primary evidence and base the main claims on task-based metrics.","section":"Section IV-C, 'Scale-Invariant Signal-to-Distortion Ratio (SI–SDR)'"},{"comment":"The manual ridge annotations used as ground truth were 'initially estimated with [16] and subsequently manually refined post hoc,' and [16] is the same multicomponent tracker that is used for training-set synthesis and for evaluating the denoised outputs. Since the network is trained to approximate targets derived from that tracker, improved agreement with these annotations may partly reflect the denoiser conforming to the tracker's own bias rather than to independent ground truth. The manual refinement reduces but does not eliminate this concern. The authors should report the level of agreement between the raw [16] tracks and the refined annotations, or provide an evaluation against independently produced annotations or synthetic ground truth, before the ridge-tracking improvements can be taken at face value. The statement in Section V-D that this evaluation is based on only 193 USVs further limits the strength of the result.","section":"Section IV-C, 'Accuracy of Tracked Ridges'"},{"comment":"The classification results are the least entangled evidence in the paper, but they are presented in a way that does not fully support the claim that the denoiser improves classification. As the manuscript itself states, the experiments evaluate the integrated pipeline of denoising plus classifier retraining, and a cross-condition test (e.g., a classifier trained on original audio and evaluated on denoised audio, or vice versa) would be needed to attribute gains to the denoiser alone. Moreover, no confidence intervals, seed variance, or significance tests are reported for the macro-F1 differences in Tables IV and V (89% vs 83%, 72% vs 69%, 61% vs 57%), and the out-of-sample improvements are small. The authors should provide uncertainty estimates and, if they wish to claim a denoising benefit for classification, run the cross-condition experiment they identify as beyond the scope of the present study.","section":"Section V-C and Section IV-C final paragraph"}],"minor_comments":[{"comment":"The conclusion contains malformed numeric expressions: '1.4 ±0.5 3.5 kHz' and '0.8 ±0.5 1.5 kHz' should be '1.4 ± 3.5 kHz' and '0.8 ± 1.5 kHz', respectively.","section":"Section VI"},{"comment":"The table entries '9766 0.8±0.3' and '93830.9±0.2' appear to be missing separators; they should presumably read '97 66' and '93 83' for precision and recall.","section":"Table III"},{"comment":"The description of the Biodenoising evaluation ('we rewrote recordings at 16 kHz without waveform resampling') is technically unclear and should be explained more precisely, because the sample-rate change affects the mapping of ultrasonic content onto the model's input range.","section":"Section IV-C, Biodenoising baseline"},{"comment":"The SI-SDR comparison shows error bars but no statistical tests across methods or SNR levels; adding paired significance tests or confidence intervals would strengthen the comparison against baselines.","section":"Figure 3"},{"comment":"The abstract says the method enhances ridge tracking 'compared to our previous signal-processing approach,' but the experiments compare against several baselines (DeepSqueak, noisereduce, Biodenoising, and the analytical denoiser); the wording should reflect the full comparison.","section":"Abstract"},{"comment":"The text uses both 'first harmonic partial' and 'harmonic partial' inconsistently; the authors should define whether the first harmonic is the second frequency component (2f0) or the fundamental itself, and use consistent terminology throughout.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The central issue is that the two most striking quantitative results (SI-SDR and ridge-tracking accuracy) are generated by or annotated with the same ridge-tracking pipeline that produces the training targets, which is a self-referential evaluation loop. The classification results are more independent but lack error bars and a cross-condition ablation. I would ask for the additional experiments described in the major comments before considering the paper for acceptance; the authors' disclosure of code and data is a positive factor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this is a credible, well-scoped applied paper that introduces a genuinely transferable trick—synthesize clean training targets from automatically tracked ridges, then train a U-Net with a ridge-weighted loss. The authors shipped code, data, and manual annotations, and the ablation on loss weighting is informative. But the two headline numbers (SI-SDR from -10.17 to 16.52 dB, ridge deviation from 1.4±3.4 to 0.8±1.5 kHz) are less impressive than they look, because both are measured against references derived from the same ridge tracker used to build the training targets. The cleanest evidence is the classification macro-F1 on the out-of-sample domesticated strain (57% to 61%)—modest but genuinely out-of-sample.\n\nWhat's new: unlike Biodenoising, which uses a pretrained speech enhancer to make pseudo-clean targets, this generates targets analytically from the ridge parameters themselves. That's a nice idea, and the paper is honest that it depends on ridges being 'moderately accurate.' The ablation shows the ridge-weighted loss matters: removing it drops harmonic recall from 83% to 66%, which is a concrete, non-obvious finding.\n\nWhere I'd push back: the SI-SDR section is self-referential and the paper knows it—it calls the test 'a controlled, relative comparison'—but then the conclusion quotes the dB numbers as if they were absolute. I'd move that to a sanity-check framing. The ridge-tracking evaluation is similarly entangled: the manual annotations were seeded by [16] and then refined, and [16] is the same tracker used in the synthesis pipeline, so the gains partly reflect the network learning the tracker's bias. The classification evidence is better, but lacks error bars or seed variance, and the 'out-of-sample' wild subset comes from the same corpus, so the domesticated set is the only fully independent test.\n\nAlso, Eq. (5) never defines b, the half-width of the ridge-weighting band, and that's needed to reproduce the central loss. Minor, but fixable. The paper's own limitations section covers faint calls and the 193-USV annotation set, so I'm not raising those as new flaws.\n\nOverall: the core method holds up for practical purposes. I'd send it to a serious referee, with the request that the synthetic metrics be reframed and the hyperparameters completed. Worth a reading-group slot if anyone works on bioacoustic signal processing.","headline":"Ridge-based training-set synthesis is a sound, useful bioacoustic denoising trick, but the SI-SDR and ridge-tracking headlines are partly self-referential; trust the modest classification gains.","tokens_in":25156,"tokens_out":3192,"would_cite":true,"duration_ms":30577,"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":"A supervised denoiser for bioacoustic recordings can be trained without clean audio by synthesizing clean targets from automatically tracked frequency ridges, and on mouse ultrasonic vocalizations this improves denoising, ridge tracking…","keywords":["bioacoustic denoising","training set synthesis","ridge tracking","ultrasonic vocalizations","complex ratio mask","U-Net","ridge-guided loss","mouse USVs"],"falsifier":"Take a synthetic vocalization with a known clean waveform and known ridge and harmonic structure, add real field noise, run the full pipeline—track, mask, synthesize target, train, denoise—and measure whether the denoised output recovers spectral energy that lies outside the 5-bin Gaussian mask but is part of the true synthetic signal; systematic loss of that energy would show the targets are teaching the network to delete genuine signal content.","tokens_in":23816,"feed_emoji":"🐭","tokens_out":8505,"duration_ms":73351,"temperature":0.7,"pith_summary":"This paper aims to solve a chicken-and-egg problem in bioacoustics: supervised denoising needs clean training audio, but clean audio is exactly what field recordings of weakly vocalizing animals do not have. The authors propose to synthesize the clean training signal from the noisy recording itself, using automatically tracked ridges—the time-frequency contours of the fundamental frequency and its first harmonic—as a proxy for the true vocalization. A Gaussian mask around those ridges selects the signal-dominated regions of a reassigned spectrogram, and the masked representation is inverted into audio that serves as the training target. A U-Net is then trained to predict a complex ratio mask from noisy spectrograms, with a ridge-guided loss that up-weights ridge regions and penalizes residual noise elsewhere. In mouse ultrasonic vocalizations this pipeline improves signal-to-distortion ratio, reduces fundamental-ridge tracking error on field recordings from 1.4 ± 3.4 kHz to 0.8 ± 1.5 kHz, and raises syllable classification F1 on denoised-and-retrained data.","feed_headline":"Mouse-call denoiser trains on ridges, no clean audio needed","feed_subtitle":"Fundamental-frequency tracking error drops from 1.4 to 0.8 kHz; classifier F1 rises from 83 to 89 percent.","key_machinery":"The load-bearing object is the ridge-guided synthesis-from-reassignment pipeline. A ridge is a frequency contour—a curve of local maxima in a time-frequency representation—that here tracks the fundamental frequency and the first harmonic partial of each mouse ultrasonic vocalization. The pipeline produces pseudo-clean training audio by masking the reassigned spectrogram with a truncated Gaussian centered on those ridges and inverting the masked representation through convolution with the Hann-window kernel and fast Griffin-Lim phase retrieval, and it reuses the same ridges to build weight matrices for the loss, emphasizing ridge-adjacent time-frequency bins and penalizing energy elsewhere. The denoiser itself is a U-Net predicting a complex ratio mask—a time-frequency mask applied to the complex spectrogram so that both magnitude and phase are corrected. The argument is carried by the tandem use of ridges as both target generator and loss guide: that is what lets a supervised denoiser be trained from noisy field recordings alone.","core_discovery":"The paper's central claim is that imperfect, automatically tracked ridges can replace clean recordings as supervised training targets for bioacoustic denoising. For each noisy mouse ultrasonic vocalization, the authors track the fundamental-frequency ridge and the first harmonic partial using multitaper reassigned spectrograms, build a truncated Gaussian mask around both ridges (window length L=5 bins), apply this mask to the reassigned spectrogram, and invert the masked representation—via convolution with the analysis window and fast Griffin-Lim phase retrieval—into a synthetic 'clean' signal. That synthetic signal, paired with noise added back at controlled SNRs, trains a U-Net to predict a complex ratio mask; a ridge-guided loss up-weights the ridge neighborhoods and adds a penalty on non-ridge energy. The paper reports that this trained denoiser raises mean SI-SDR of synthetic test signals from -10.17 dB to 16.52 dB at the lowest input SNR, cuts fundamental-ridge frequency deviation on field recordings from 1.4 ± 3.4 kHz to 0.8 ± 1.5 kHz, and, when classifiers are retrained on denoised data, raises macro-F1 from 83% to 89% on in-sample wild-mouse data and from 69% to 72% on out-of-sample noisy wild-mouse data.","pith_inferences":["Because the synthetic test set is generated by the same ridge-mask-inversion pipeline that creates the training targets, the reported SI-SDR numbers measure self-consistency of the synthesis as much as true recovery; an independent check with real clean recordings of a captive or laboratory animal would separate the two.","The narrow 5-bin Gaussian mask and the restriction to fundamental plus first harmonic mean that any genuine signal energy outside those contours is omitted from the targets; species with rich harmonic stacks or strong nonlinear frequency modulations would be a natural stress test for the approach.","The paper filters training to the 83.7% of USVs whose ridges are judged correct, so an iterative loop—denoise, retrack, re-synthesize, retrain—is a natural extension that could recover some of the 13.7% incorrect-ridge cases.","Preserving multiple ridge trajectories in the output hints at a source-separation use: overlapping calls from two mice, currently excluded as invalid samples, might be disentangled by treating each ridge as a separate synthesis target."],"forward_implications":["Any tonal bioacoustic signal with trackable ridges—marmoset calls, dolphin whistles, bat calls—could be denoised with this recipe, since the ridge tracker is swappable and no clean recordings are required.","Retraining downstream classifiers on denoised audio yields consistent gains: macro-F1 rises from 83% to 89% on in-sample wild-mouse USVs, from 69% to 72% on out-of-sample noisy wild-mouse USVs, and from 57% to 61% across a domesticated mouse strain.","Ridge tracking itself improves after denoising: fundamental-ridge frequency deviation falls from 1.4 ± 3.4 kHz to 0.8 ± 1.5 kHz, and harmonic-partial recall rises from 58% to 83% relative to the previous signal-processing pipeline.","The ridge-guided loss is doing real work: removing it keeps fundamental-ridge precision at 95–96% but raises frequency deviation and lowers harmonic recall, so the weighting is what preserves faint components.","Because the synthesis targets are analytic—built from explicit amplitude, instantaneous frequency, and phase parameters—the training data are interpretable and controllable, unlike targets coming from a pretrained denoiser."],"supporting_citations":[{"why":"Supplies the multicomponent ridge-tracking method (multitaper reassigned STFT/CQT plus autocorrelation for harmonics) that yields the frequency contours used for synthesis, and provides the baseline ridge-tracking numbers.","marker":"[16]"},{"why":"Contributes the reassigned-spectrogram inversion procedure (convolution with the window kernel plus fast Griffin-Lim phase retrieval) that the paper adapts with a ridge-guided mask to synthesize pseudo-clean targets.","marker":"[54]"},{"why":"Defines the complex ratio mask that the U-Net predicts, so that denoising corrects phase as well as magnitude.","marker":"[55]"},{"why":"Provides the U-Net encoder-decoder with skip connections used as the denoiser architecture.","marker":"[44]"},{"why":"Supplies a commonly used USV ridge-tracker whose performance is a comparison baseline for ridge-tracking accuracy.","marker":"[62]"},{"why":"Provides the BootSnap CNN classifier and its domesticated-strain test set, used to evaluate the integrated denoising-and-retraining pipeline.","marker":"[64]"},{"why":"Supplies the noisereduce spectral-gating baseline compared on SI-SDR in stationary and non-stationary modes.","marker":"[26]"},{"why":"Supplies the Biodenoising pseudo-clean-target baseline, the alternative unsupervised denoising approach the paper contrasts with its analytic ridge synthesis.","marker":"[46]"},{"why":"Provides the wild house-mouse USV recordings and their syllable-type labels that form the case-study dataset.","marker":"[17]"},{"why":"Supplies the fast Griffin-Lim algorithm used for phase retrieval when synthesizing the clean training target.","marker":"[60]"}],"fun_headline_variants":["Train bioacoustic denoiser on ridge-derived masks, not clean audio","Mouse-call denoiser learns from ridge tracks, not clean clips","Ridge-synthesized training data boosts mouse-call denoising without clean audio","No clean audio? Train on ridge-tracking masks for mouse call denoising","Ridge-tracked synthetic training cuts mouse-call noise without clean data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the automatically tracked ridges on noisy recordings are accurate enough that a clean training signal reconstructed from a narrow Gaussian mask around those ridges contains the true vocalization; if a ridge is wrong or a harmonic is missed, the target excludes real signal energy and the denoiser learns to erase it.","fun_headline_variants_meta":{"raw":{"variants":["Train bioacoustic denoiser on ridge-derived masks, not clean audio","Mouse-call denoiser learns from ridge tracks, not clean clips","Ridge-synthesized training data boosts mouse-call denoising without clean audio","No clean audio? Train on ridge-tracking masks for mouse call denoising","Ridge-tracked synthetic training cuts mouse-call noise without clean data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001002,"raw_usage":{"total_tokens":4337,"prompt_tokens":1137,"completion_tokens":3200,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":753,"completion_tokens_details":{"reasoning_tokens":3098}},"tokens_in":753,"tokens_out":3200,"duration_ms":21679,"temperature":1.0,"reasoning_tokens":3098,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:14:15.387197+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic vocalization with a known clean waveform and known ridge and harmonic structure, add real field noise, run the full pipeline—track, mask, synthesize target, train, denoise—and measure whether the denoised output recovers spectral energy that lies outside the 5-bin Gaussian mask but is part of the true synthetic signal; systematic loss of that energy would show the targets are teaching the network to delete genuine signal content.","supporting_citations":[{"cited_title":"Robust multicomponent tracking of ultrasonic vocalizations,","cited_arxiv_id":null,"evidence_quote":"Supplies the multicomponent ridge-tracking method (multitaper reassigned STFT/CQT plus autocorrelation for harmonics) that yields the frequency contours used for synthesis, and provides the baseline ridge-tracking numbers."},{"cited_title":"Biocppnet: automatic bioacoustic source separation with deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Contributes the reassigned-spectrogram inversion procedure (convolution with the window kernel plus fast Griffin-Lim phase retrieval) that the paper adapts with a ridge-guided mask to synthesize pseudo-clean targets."},{"cited_title":"Rajbamshi, P","cited_arxiv_id":null,"evidence_quote":"Defines the complex ratio mask that the U-Net predicts, so that denoising corrects phase as well as magnitude."},{"cited_title":"Automatic mouse ultrasound detector (A-MUD): A new tool for processing rodent vocalizations,","cited_arxiv_id":null,"evidence_quote":"Supplies a commonly used USV ridge-tracker whose performance is a comparison baseline for ridge-tracking accuracy."},{"cited_title":"Female presence and estrous state influence mouse ultrasonic courtship vocalizations,","cited_arxiv_id":null,"evidence_quote":"Provides the BootSnap CNN classifier and its domesticated-strain test set, used to evaluate the integrated denoising-and-retraining pipeline."},{"cited_title":"Courtship vocalizations of wild house mice show highly dynamic changes and correlate with male copulatory success,","cited_arxiv_id":null,"evidence_quote":"Provides the wild house-mouse USV recordings and their syllable-type labels that form the case-study dataset."},{"cited_title":"Reassignment and synchrosqueezing for general time–frequency filter banks, subsampling and processing,","cited_arxiv_id":null,"evidence_quote":"Supplies the fast Griffin-Lim algorithm used for phase retrieval when synthesizing the clean training target."}],"review_version":1}