{"id":"86768a85-b62d-48fd-886a-08c64519ed2a","arxiv_id":"2412.17851","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Noisereduce, a no-training spectral gating method, outperforms classical noise reduction baselines across speech, bioacoustics, neurophysiology, and seismology, and is a fast, domain-general baseline.","lead":"This paper validates Noisereduce, a lightweight, training-free spectral gating algorithm, on speech, birdsong, brain recordings, and seismic signals, reporting that it beats classic noise reduction methods and runs fast. It also introduces a new public benchmark for birdsong noise reduction.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bioacoustics comparison gives Noisereduce a matched noise clip while baselines are blind; the reported advantage may be an artifact of this asymmetry.","rationale":"The reader's weakest_assumption concerns the general validity of the spectral threshold (mean + k·std) for separating signal and noise—a theoretical limitation. My concern is a concrete experimental design flaw: in the bioacoustics benchmark, Noisereduce is given a noise-only clip drawn from the same noise source and SNR as the added noise, while the comparison methods (Wiener, Savitzky-Golay) are not. This is a load-bearing issue because the paper's central claim is comparative (\"outperforms other conventional algorithms\") and domain-general. If the advantage in one domain disappears when the oracle noise clip is removed or when baselines are given equal noise information, the headline claim is overstated. This is not merely a missing significance test or a vague exclusion criterion; it directly undermines the fairness of the comparison. The concern is testable: the code and dataset are publicly available, and re-running the experiment in the two conditions described would settle whether the asymmetry explains the result. I therefore recommend keeping the paper conditional: it should be published only if the authors either confirm that the baseline methods were given equivalent noise information (which the text does not support) or revise the bioacoustics evaluation to be fair and re-report the results. This does not require rejection, because the speech, electrophysiology, and seismology results may still support the algorithm's usefulness, and the Birdsong dataset is a valuable resource.","tokens_in":17290,"tokens_out":6142,"duration_ms":53800,"concrete_test":"Re-run the Birdsong NOIZEUS experiments (Tables 5 and 6) under two conditions: (A) Noisereduce in non-stationary mode with no noise clip, and (B) give Wiener and Savitzky-Golay the same matched noise clip (e.g., use it to estimate the noise PSD for Wiener and spectral subtraction) while keeping Noisereduce stationary. If Noisereduce no longer dominates SegSNR and SDR in either condition, its reported advantage is an artifact of the oracle noise clip rather than a general algorithmic superiority.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The Birdsong NOIZEUS dataset (Sec. 4.3) explicitly includes, for each song/noise pair, a 1-second clip of noise sampled from the same soundscape at the same SNR (\"for each song and noise clip we included a 1-second clip of noise sampled randomly at the same SNR of the audioclip\"). The algorithm description (Sec. 1.1, Eqs. 1–5) shows that stationary Noisereduce computes its threshold from an optional noise clip Xnoise; without one, it uses the signal itself, which would be inappropriate for continuous birdsong. The experiments in Sec. 2.2 compare Noisereduce against Wiener and Savitzky-Golay, neither of which is given a noise clip—both must estimate noise from the corrupted signal. Noisereduce therefore receives near-oracle noise statistics (pure noise from the exact source and SNR), while the baselines are blind to this information. This asymmetry makes the reported \"Noisereduce outperforms the other conventional algorithms\" in SegSNR and SDR unsurprising and does not test the algorithm's behavior when such a matched noise clip is unavailable, which is the common real-world scenario. The central claim of domain-general superiority is thus not supported by the bioacoustics experiment as run.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Noisereduce, an open-source spectral gating algorithm for time-series noise reduction, and evaluates it on four domains: speech, bioacoustics, electrophysiology, and seismology. The algorithm computes a per-frequency threshold from noise statistics, builds a binary time-frequency mask, optionally smooths it, and applies it to the signal STFT; a non-stationary variant computes the threshold from sliding windows on the signal itself. The authors compare Noisereduce against several conventional algorithms (Wiener, Iterative Wiener, Savitzky-Golay, Spectral Subtraction, Subspace) and, in speech and seismology, against deep-learning baselines. They report that Noisereduce outperforms the conventional algorithms on the tested metrics and runs faster than the deep-learning baselines, supporting its proposed role as a domain-general tool and a baseline for machine-learning methods.","tokens_in":17556,"tokens_out":6167,"duration_ms":54258,"significance":"If the claims hold, the paper provides a useful validation of a widely used open-source tool, with concrete benefits: the algorithm requires no training data, has few parameters, supports GPU acceleration, and the authors release the benchmark data and reproduction scripts. The head-to-head comparisons with Denoiser and DeepDenoiser are informative for practitioners choosing between conventional and learned methods. However, the strength of the central claim depends on the correctness of the algorithm description and on the fairness of the comparisons; both currently have load-bearing issues that require correction before the reported results can be taken as supporting the advertised conclusions.","major_comments":[{"comment":"Eq. (14) defines the masked STFT as Sm = conv2d(SX, SF), where SF is the smoothing filter, but the accompanying text says the mask Msmooth is applied to SX. As written, the mask M never enters the output and the operation would only smooth the spectrogram, not perform noise gating. This contradicts Step 2.4 in Section 1.1 and the block diagram in Fig. 1, where the smoothed mask is multiplied elementwise with the signal STFT. Please correct Eq. (14) to an elementwise multiplication of Msmooth with SX, and adjust Eqs. (9)–(13) so that they describe smoothing of the mask rather than of the signal spectrogram.","section":"4.2.1, Eq. (14)"},{"comment":"The bioacoustics comparison appears to give Noisereduce an information advantage over the baselines. Section 4.3 states that for each song/noise pair the Birdsong NOIZEUS dataset includes a one-second noise clip at the same SNR, and Section 2.2 does not report whether the other algorithms received this clip. If Noisereduce was given the matched noise-only clip while Wiener and Savitzky-Golay had to estimate noise from the corrupted signal, then the large improvements in Tables 5 and 6 reflect asymmetric input information rather than algorithmic superiority. Please state explicitly which inputs each method received, and either provide the noise clip to all methods or evaluate Noisereduce without it; otherwise the claim that Noisereduce outperforms conventional algorithms in bioacoustics is not supported.","section":"2.2 / 4.3, Bioacoustics comparison"},{"comment":"The parameter prop_decrease is described as a core control for how much below-threshold noise is removed, but it is absent from the formal algorithm in Section 4.2.1. Eq. (8) defines a binary mask and Eq. (14) does not include prop_decrease at all, so a reader cannot reproduce the behavior of the published package from the equations. Please add the missing equation (e.g., a soft scaling of the mask by prop_decrease) or remove the parameter from the algorithm description and state the default mask is binary.","section":"4.2.2 / Table 9"},{"comment":"The comparative claims rest on mean differences without inferential statistics. In several rows the reported SEMs overlap (e.g., Table 7, SNR 15: Noisereduce 0.069 ± 0.019 vs Wiener 0.080 ± 0.024; Table 5, SNR 15: Savitzky-Golay 11.91 ± 2.47), and Section 2.3 reports AUC differences from a single simulated recording with no uncertainty estimate. Please report sample sizes and appropriate significance tests or paired comparisons, or temper the wording from “outperforms” to “tends to improve” where the evidence is not conclusive.","section":"2.2–2.4, Tables 5–8 and Fig. 5"}],"minor_comments":[{"comment":"It is not stated whether Noisereduce was provided with a noise-only clip in the NOIZEUS speech evaluation. This is important because the algorithm's stationary variant either uses a supplied noise clip or falls back to the signal itself; please specify which protocol was used for Tables 1–4.","section":"2.1, speech experiments"},{"comment":"The symbol W(f,t) in Eq. (13) is not defined; presumably W = Lf ⊗ Lt, but this should be stated explicitly.","section":"4.2.1, Eq. (13)"},{"comment":"The parameter names are inconsistent: the text refers to “n_std_thresh_stationary” while Table 9 lists “n_std_thresh”, and Eq. (5) uses k. Please unify the notation, and verify the default of 1 ms for “noise_window_size_nonstationary_ms”, which seems extremely short for most signals.","section":"4.2.2 / Table 9"},{"comment":"The statement that Noisereduce outperforms the baselines “across all SNR levels” is stronger than the data support at high SNR, where differences are small and confidence intervals overlap; consider adding statistical tests or softening the claim.","section":"2.4, Tables 7–8"},{"comment":"There is a typo in the caption: “Reciever Operator Characteristic” should be “Receiver Operating Characteristic.”","section":"Fig. 5 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is an engineering-style validation of a tool that is already widely used, and the released benchmark and code are valuable. The main technical risks are the incorrect formal description of the masking step and the asymmetric bioacoustics comparison; both are fixable but are central to the stated claims. I would not reject on novelty grounds, but the authors should be asked to correct Eq. (14), clarify and equalize the comparison protocol, and add the missing statistical support before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an honest validation of a widely used open-source tool, not a new method, and it deserves serious referee time. The biggest issue is a fairness problem in the bioacoustics experiment that the authors should fix.\n\nWhat is actually new and good: the paper validates Noisereduce, a spectral gating algorithm that has been public for over five years and gets real use in bioacoustics, ephys, and elsewhere. That validation was missing, and this is a reasonable first pass at it. It is honest about the algorithm's heritage, Boll 1979 is cited up front, no novelty overclaiming. The new Birdsong NOIZEUS benchmark is a real contribution, public on Zenodo, built from clean starling recordings with eight real soundscape noise types at four SNRs. Code and configs for all experiments are on GitHub. The electrophysiology and seismology evaluations use ground truth and include deep-learning baselines that the paper lets win when they should.\n\nSoft spots, in proportion. The main one is the bioacoustics comparison. The dataset includes a 1-second pure-noise clip for each song/noise pair at the same SNR, and stationary Noisereduce computes its threshold from exactly that kind of clip. Wiener and Savitzky-Golay get nothing; they estimate noise from the corrupted signal alone. So the reported SegSNR/SDR advantage in Section 2.2 is close to oracle-assisted, and the 'outperforms other conventional algorithms' claim there does not test the common case where no matched noise clip exists. I checked this reading carefully against the paper, and the concern holds. Fix: give the baselines the same noise clip, or run Noisereduce without it, and report both.\n\nSecond: Eq. 14 in the supplementary describes applying the mask as convolving the signal STFT with the smoothing filter, which contradicts the element-wise multiplication in Figure 1 and the shipped code. Typo-level, but in a methods paper a reader learning the algorithm from the text will be misled; fix it.\n\nThird, smaller: the seismic dataset description excludes 'low SNR' signals without defining the cutoff; the metric tables and runtime plot lack significance tests and error bars beyond SEM; and the non-stationary claim rests on one airplane-over-robin example, with the paper hedging in the text but the abstract saying it 'handles both stationary and non-stationary noise' more broadly than that evidence supports.\n\nThe reader's report is broadly right; I would just add that the noise-clip asymmetry matters more than the other soundness issues combined.\n\nWho this is for: anyone doing bioacoustic or general time-series denoising who needs a baseline method and a benchmark to compare against. It deserves a serious referee; my own verdict is conditional until the bioacoustics comparison is rerun fairly.","headline":"Honest validation of a widely used spectral gating tool with a genuinely useful benchmark; the bioacoustics comparison hands Noisereduce a matched noise clip the baselines never see, and that asymmetry should drive the revision.","tokens_in":18054,"tokens_out":7035,"would_cite":true,"duration_ms":48576,"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":"The paper claims Noisereduce, a spectral gating algorithm that requires no training data, outperforms other conventional noise reduction algorithms on speech, birdsong, simulated neural spikes, and seismic waveforms, making it a fast…","keywords":["noise reduction","spectral gating","time-series signals","speech enhancement","bioacoustics","electrophysiology","seismology","denoising baseline"],"falsifier":"Rerun the paper's NOIZEUS comparison with babble noise at a signal-to-noise ratio of 0 dB and measure STOI or PESQ; if Noisereduce does not exceed every conventional baseline listed (Wiener, iterative Wiener, subspace, spectral subtraction, Savitzky-Golay), the paper's main speech superiority claim is false. A similar test on non-stationary noise with abrupt amplitude changes would probe whether the sliding-window variant truly adapts faster than the stationary version.","tokens_in":17087,"feed_emoji":"🎛️","tokens_out":8748,"duration_ms":70727,"temperature":0.7,"pith_summary":"The paper sets out to establish that Noisereduce, a spectral gating algorithm, is a fast, domain-general noise reduction method that works across speech, bioacoustics, electrophysiology, and seismology without needing training data. It argues this matters because many real-world signal domains lack the large labeled datasets that deep-learning denoisers require, so a lightweight conventional method can serve both as a practical tool and as a fair baseline for judging machine-learning models. The paper reports that Noisereduce outperforms other conventional algorithms (Wiener, iterative Wiener, subspace, spectral subtraction, and Savitzky-Golay) at every tested SNR on the NOIZEUS speech corpus, and also leads on a new birdsong benchmark, simulated spike detection, and seismic event onset detection. Its non-stationary variant, which recomputes noise statistics in a sliding window, is shown to handle noise that drifts over time, such as an airplane passing over a bird recording.","feed_headline":"Noisereduce beats conventional filters in four signal domains","feed_subtitle":"A training-free spectral gating algorithm cleans speech, birdsong, neural spikes, and seismograms.","key_machinery":"The load-bearing object is the per-frequency noise threshold $\\mathrm{thresh}_n(f)=\\mu_n(f)+k\\sigma_n(f)$ computed from the decibel spectrogram of a noise estimate, which turns into a binary spectral mask $M(f,t)$. The mask is the whole algorithm: it decides, independently for each time-frequency cell, whether the energy is signal (keep) or noise (zero it), with an optional triangular smoothing kernel over frequency and time to reduce artifacts. Everything else, including the STFT, the inverse STFT, and the sliding-window statistics in the non-stationary variant, exists to feed this threshold. The argument stands or falls on whether such an amplitude threshold, set from noise statistics, separates signal from noise across domains.","core_discovery":"Noisereduce's central claim is that a simple per-frequency amplitude threshold cleanly separates signal from noise in a wide range of time-series recordings. The algorithm computes the short-time Fourier transform of a noise clip, converts magnitudes to decibels, and for each frequency $f$ sets $\\mathrm{thresh}_n(f) = \\mu_n(f) + k\\sigma_n(f)$, where $\\mu_n$ and $\\sigma_n$ are the mean and standard deviation of the noise spectrogram across time and $k$ is a sensitivity hyperparameter. It then forms a binary mask $M(f,t)=1$ when the signal spectrogram exceeds the threshold and $0$ otherwise, optionally smooths the mask in time and frequency, multiplies the signal STFT by the mask, and inverts back to the time domain. The non-stationary variant computes the same statistics over a sliding window on the signal itself, so the threshold adapts as the noise floor changes. The paper argues that this training-free mask is fast, requires minimal computational resources, and outperforms other conventional algorithms in all four tested domains, while remaining competitive with deep-learning models at a fraction of the cost.","pith_inferences":["Editorial inference: The paper's domain-generality claim implies Noisereduce should transfer to other time-series domains with spectro-temporal structure, such as ECG artifact removal or sonar, as long as a noise-only clip or a slowly drifting noise floor is available; this transfer is not tested in the paper.","Editorial inference: Because the mask is binary and amplitude-based, the non-stationary variant will lag abrupt noise-floor changes and will erase low-amplitude signal that overlaps the noise distribution; an adaptive or soft mask, which the paper mentions as future work, is the natural next test.","Editorial inference: The authors' caveat that Noisereduce should be used only for spike detection, not clustering, suggests that any reconstruction-based downstream use, such as seismic waveform inversion or acoustic feature extraction, should validate that the mask does not distort the features of interest.","Editorial inference: The Birdsong NOIZEUS dataset could become a reusable benchmark for bioacoustic denoising, enabling future machine-learning methods to be compared fairly across studies; the authors created it but do not make this claim explicitly."],"forward_implications":["In speech, Noisereduce can be adopted as the conventional baseline for comparing new machine-learning enhancers, since it beats Wiener, iterative Wiener, subspace, spectral subtraction, and Savitzky-Golay at 0, 5, 10, and 15 dB SNR on NOIZEUS.","In bioacoustics, the new Birdsong NOIZEUS benchmark gives researchers a standard dataset, and Noisereduce provides a strong first-pass denoiser for recordings where no labeled training data exist.","In electrophysiology, Noisereduce improves automated spike detection (ROC AUC 0.97 versus 0.91 unfiltered, 0.94 Wiener, and 0.96 Savitzky-Golay) and can be used as a preprocessing step before spike sorting, though not for clustering because it alters spike shapes.","In seismology, Noisereduce reduces STA/LTA onset-detection error more than Wiener or Savitzky-Golay at low SNR for both white and pink noise, and comes closer to DeepDenoiser's performance than other conventional methods.","GPU-accelerated Noisereduce runs fast enough for real-time or near-real-time use, making the algorithm practical on large datasets and resource-limited systems where deep denoisers are too slow."],"supporting_citations":[{"why":"Defines spectral gating and subtraction, the approach Noisereduce generalizes.","marker":"[5]"},{"why":"Supplies the classic spectral subtraction algorithm used as a speech comparison baseline.","marker":"[6]"},{"why":"Provides the Wiener and iterative Wiener filters used as comparison baselines.","marker":"[25]"},{"why":"Provides the Savitzky-Golay filter used as a comparison baseline across all domains.","marker":"[47]"},{"why":"Supplies the NOIZEUS noisy speech corpus used for speech enhancement evaluations.","marker":"[29]"},{"why":"Supplies the Denoiser deep-learning model that sets the machine-learning comparison point in speech.","marker":"[10]"},{"why":"Supplies the DeepDenoiser model and seismic denoising setup used for seismology comparisons.","marker":"[59]"},{"why":"Generates the simulated extracellular recordings with ground-truth spikes used in the electrophysiology test.","marker":"[7]"},{"why":"Provides the spike detection algorithm used for the ROC and AUC evaluation.","marker":"[8]"},{"why":"Introduces the Birdsong NOIZEUS benchmark dataset used for bioacoustic evaluation.","marker":"[44]"}],"fun_headline_variants":["Noisereduce: one training-free filter for all noise types","Spectral gating algorithm cleans four time-series domains","Noisereduce removes noise without training data","Fast noise reduction for speech, birdsong, neural, seismic","Noisereduce: domain-general noise reduction in 4 domains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the noise measured in a quiet clip, or in short windows around the signal, is the same noise that corrupts the signal, and that in each frequency band the signal's loudness is higher than the noise's average loudness plus a few fluctuations; when the noise shifts or the signal is faint, the mask erases signal or leaves noise.","fun_headline_variants_meta":{"raw":{"variants":["Noisereduce: one training-free filter for all noise types","Spectral gating algorithm cleans four time-series domains","Noisereduce removes noise without training data","Fast noise reduction for speech, birdsong, neural, seismic","Noisereduce: domain-general noise reduction in 4 domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1177,"prompt_tokens":887,"completion_tokens":290,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":207}},"tokens_in":503,"tokens_out":290,"duration_ms":2381,"temperature":1.0,"reasoning_tokens":207,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:38:01.165845+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the paper's NOIZEUS comparison with babble noise at a signal-to-noise ratio of 0 dB and measure STOI or PESQ; if Noisereduce does not exceed every conventional baseline listed (Wiener, iterative Wiener, subspace, spectral subtraction, Savitzky-Golay), the paper's main speech superiority claim is false. A similar test on non-stationary noise with abrupt amplitude changes would probe whether the sliding-window variant truly adapts faster than the stationary version.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines spectral gating and subtraction, the approach Noisereduce generalizes."},{"cited_title":"Savitzky and M","cited_arxiv_id":null,"evidence_quote":"Provides the Savitzky-Golay filter used as a comparison baseline across all domains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the NOIZEUS noisy speech corpus used for speech enhancement evaluations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Generates the simulated extracellular recordings with ground-truth spikes used in the electrophysiology test."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the spike detection algorithm used for the ROC and AUC evaluation."},{"cited_title":"Sainburg and A","cited_arxiv_id":null,"evidence_quote":"Introduces the Birdsong NOIZEUS benchmark dataset used for bioacoustic evaluation."}],"review_version":1}