{"id":"1b106f89-56e8-47af-b7d6-4cd8d8a9df1c","arxiv_id":"2507.01821","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"WindNetLite, a 249K-parameter dual-encoder ULCNet variant, achieves wind noise reduction performance comparable to ULCNet at roughly 40% of its computational cost.","lead":"This paper introduces WindNetLite, a small neural network for removing wind noise from audio recordings that is designed to run in real time on embedded devices. The model gives more processing power to low frequencies, where wind noise lives, and it matches larger models in quality while running about two and a half times cheaper.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"WindNetLite's complexity savings rely on an unvalidated claim that wind energy above 4 kHz is negligible; if high-frequency wind occurs, the reduced HF encoder will fail to suppress it, breaking the comparable-performance claim.","rationale":"The reader's weakest assumption identifies exactly the spectral concentration assumption, and I agree that it is load-bearing. WindNetLite's complexity advantage comes from offloading most of the modeling burden to the LF encoder; the HF encoder is deliberately starved. If the wind signal contains substantial energy above 4 kHz, the HF path cannot learn to suppress it, and the model would either pass wind through or distort the desired signal while attempting to remove it. The paper never provides the spectral statistics of its own 10-hour wind database, nor does it compare against wind with high-frequency content. Since the central claim is that WindNetLite is 'suitable for embedded and mobile audio applications' in general, this unverified physical assumption is a real soft spot. The proposed test (controlled HF wind energy) would directly falsify or confirm the assumption's impact. I therefore keep the reader's CONDITIONAL verdict: the engineering is plausible, but the generality of the result is not established until the spectral assumption is checked.","tokens_in":8277,"tokens_out":9604,"duration_ms":107437,"concrete_test":"Build a test set using the same clean speech/music signals but with wind noise spectrally shaped so that a controlled fraction (e.g., 10%) of total wind energy lies above 4 kHz. Run WindNetLite and WindNet in extraction mode (the best configuration) and compare leakage (Eq. 11) and SI-SDR to the original Table 1. If WindNetLite's SI-SDR drops by more than 1 dB relative to WindNet, or its leakage degrades by more than 0.5 dB, the low-frequency assumption is load-bearing and the general suitability claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central architectural trade-off is set by the dual-encoder split in Section 3.1. The paper states wind 'typically exhibits high energy in the low-frequency range and negligible energy at frequencies higher than 4 kHz' and assigns only the first 5 sub-bands (approximately 0-4 kHz at 16 kHz) to the high-capacity LF encoder, while the HF encoder uses average pooling, fewer filters (8/16/64 versus 32/64/96/128), and no GRU. This is the sole justification for the 249K-parameter, 73 MHz complexity budget. However, the 10-hour wind database recorded for the paper (Section 4.2) is never used to validate the 4 kHz cutoff, and no external reference is provided. Because the test sets are mixed from the same wind database, the evaluation is blind to wind with non-negligible high-frequency content. If such wind occurs in practice, the HF path has insufficient capacity to suppress it, and the claimed parity with WindNet in Table 1 would not hold outside the low-frequency wind regime. The paper also does not report the spectral statistics that could settle this. This is not an internal inconsistency, but the external validity of the central claim is load-bearing on an unverified physical assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WindNetLite, a low-complexity single-channel neural network for wind noise reduction, built by modifying the ULCNet architecture with separate low-frequency and high-frequency encoder branches. The design is motivated by the assumption that wind noise energy is concentrated below 4 kHz. The authors train and evaluate the model in two modes, termed wind rejection (direct estimation of the desired signal) and wind extraction (estimation and subtraction of the wind component), using a self-recorded 10-hour wind noise database mixed with AudioSet content and tested on VCTK speech and MUSDB18 music. They report objective metrics (PESQ, SI-SDR, PEAQ, and a leakage metric) and a MUSHRA listening test, and claim that WindNetLite, with 249K parameters and roughly 73 MHz of computational cost, achieves performance comparable to ULCNet and to the WindNet baseline while requiring only 40% of ULCNet's real-time factor and 84% of WindNet's real-time factor.","tokens_in":8513,"tokens_out":4305,"duration_ms":45366,"significance":"If the performance claims hold, the paper would make a useful contribution to real-time wind noise reduction on embedded and mobile devices. The dual-encoder frequency-aware design is a sensible way to reduce complexity, and the authors provide a concrete parameter count, real-time factor, and computational-cost estimate, as well as a newly recorded wind noise database and a public listening demo. These are valuable assets for reproducible comparison in a field where such details are often missing. However, the central claim of parity with ULCNet and WindNet is not yet fully established, because the evaluation contains methodological gaps: hyperparameter selection is not separated from test data, the ULCNet comparison is not trained on the same task, and no statistical significance or variance information is provided for the reported metrics.","major_comments":[{"comment":"The power-law compression factor alpha is selected per mode as the value that minimizes the leakage metric in Figure 3, and the same value is then used in all experiments reported in Table 1. The manuscript does not describe a held-out validation split for this selection; if the optimization was performed on the test set, the reported results are optimistically biased. Please describe the validation procedure, or re-run the comparison with a fixed alpha or with a nested validation protocol, and state which results are affected.","section":"Section 5.1 and Figure 3"},{"comment":"The comparison to ULCNet is not apples-to-apples: the ULCNet row in Table 1 comes from a model trained for general speech enhancement, not for wind noise reduction, as the authors themselves note in Section 5.2. Since the abstract and conclusions claim parity with ULCNet, this comparison should be based on a ULCNet model retrained on the same wind-noise training data, or the claim should be restricted to the WindNet baseline. As written, the ULCNet comparison does not support the central claim.","section":"Section 4.2 and Table 1"},{"comment":"All objective metrics are reported as single point estimates without error bars, confidence intervals, or significance tests. Several differences that the text interprets as comparable or superior are small (e.g., PESQ 2.26 vs. 2.23, SI-SDR 12.71 vs. 12.27 in rejection mode). The MUSHRA test in Section 5.3 uses only 10 listeners and reports no statistical analysis. To support the claim of comparable performance, please report the variance across test items or training runs and apply appropriate significance testing.","section":"Section 4.4 and Table 1"},{"comment":"The dual-encoder design and the resulting complexity savings rest on the assumption that wind noise has negligible energy above 4 kHz. Although the paper recorded a 10-hour wind database, it does not report the spectral statistics of that database to validate this cutoff. If real wind or microphone scenarios contain non-negligible high-frequency energy, the lightweight high-frequency encoder may be unable to suppress it, and the performance parity shown in Table 1 would not generalize. Please provide spectral evidence from the recorded wind data (e.g., average power spectra, energy ratio below and above 4 kHz) and, ideally, an evaluation on wind segments with stronger high-frequency content.","section":"Section 3.1 and Section 4.2"}],"minor_comments":[{"comment":"The split into rXc-low and rXc-high is described in words but never defined with dimensions; please specify the exact shapes and sub-band indices of the two tensors.","section":"Section 3.1, around Eq. (7)"},{"comment":"The SNR set in the text is typeset with stray characters: \"t´20,dB,´10,dB,0,dB,10,dB,20,dBu\" should read {-20, -10, 0, 10, 20} dB.","section":"Section 4.2"},{"comment":"The ULCNet row has no PEAQ or leakage entries for the instrumental music test set; please either report those values or state explicitly why they are omitted.","section":"Table 1"},{"comment":"The caption \"limited to the 2´ 3 kHz frequency range\" is ambiguous: it is not clear whether the wind spectrogram is shown only for the 2-3 kHz band or whether the authors claim wind energy is confined to that band, which would conflict with the text in Section 3.1. Please rephrase.","section":"Figure 2 caption"},{"comment":"The test is called a MUSHRA test, but the description does not mention the use of the ITU-R BS.1534 anchors (low anchor, hidden reference). If anchors were not used, the method should be described as a graded listening test rather than MUSHRA.","section":"Section 5.3"},{"comment":"The unit \"MHz\" is used for computational cost; if this refers to mega operations per second or clock cycles, please state the definition explicitly to avoid confusion with clock frequency.","section":"Table 1 and Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is topical and the engineering results are plausible, but the evaluation needs strengthening before the claims can be accepted. In particular, separating hyperparameter selection from test data and providing a properly trained ULCNet baseline would substantially increase confidence. If the authors are willing to release the recorded wind database (or a processed subset), that would be a valuable contribution to the community and would help verify the robustness of the 4 kHz cutoff assumption."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on arXiv:2507.01821. This is a credible engineering paper: WindNetLite adapts ULCNet to wind noise reduction with a dual-encoder split that spends capacity on low frequencies, gets 249K params / 73 MHz, and roughly matches WindNet (the ULCNet variant retrained for WNR) on speech and music test sets. The wind extraction mode actually beats WindNet on SI-SDR for both speech and music. The RTF reduction from 0.061 to 0.051 is modest but real. For embedded WNR, this is useful.\n\nWhat's genuinely new: the LF/HF dual-encoder design for WNR, and the first WNR application of the ULCNet family. The paper also records 10 hours of real wind noise with 20 mics, which is more than many WNR papers do. The comparison to ULCNet as an external benchmark is independent, though ULCNet was trained for general SE, not WNR, so it's a weak external benchmark. The real baseline is WindNet, which they constructed fairly.\n\nSoft spots, in order of importance.\n\n1. The power-law factor α is chosen per mode (0.3 rejection, 1.0 extraction) based on Figure 3, and the same test sets are then used for Table 1. No held-out split is described. This risks overfitting the test set. A simple fix: select α on a validation split and report test results. This should have been done.\n\n2. The 4 kHz cutoff that justifies the LF/HF split is never validated on their recorded wind database. They show one spectrogram with wind energy below 3 kHz, but they don't report spectral statistics across the 10-hour corpus. If some wind conditions have non-negligible high-frequency energy, the reduced HF encoder could miss it. This is an external validity concern, not an internal error. The performance claim may still hold for typical low-freq wind, but the paper should show the distribution or limit the claim.\n\n3. No error bars on objective metrics, and the MUSHRA test is 10 listeners on 10 samples with no significance testing. That's enough for a spot check, not for a strong subjective claim. The paper's informal listening caveat is honest, but the figure is presented without statistical context.\n\n4. No code or wind data released. For reproducibility, that's a gap, though not unusual for industrial-affiliated papers.\n\nThe citation pattern looks fine. The self-citation to ULCNet is appropriate; the prior WNR work is covered.\n\nOverall: the central claim is plausible and the engineering is sound. The paper deserves a serious referee, but the authors should be asked to fix the alpha selection, add error bars, and provide spectral stats or at least temper the generalization claim. Reader's CONDITIONAL verdict is fair; the stress-test's 'breaks' wording is too strong—the claim doesn't break on its own terms, but its external reach is unverified.\n\nWould I cite it? Likely, as a data point for low-complexity WNR. Reading group: maybe, if the group likes discussing evaluation pitfalls. Send to review, yes.","headline":"Solid low-complexity WNR engineering with a dual-encoder split; plausible results but evaluation has a few fixable soft spots.","tokens_in":9074,"tokens_out":3165,"would_cite":true,"duration_ms":32499,"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":"WindNetLite, with 249K parameters and 73 MHz, matches ULCNet-quality wind noise reduction at a fraction of the compute by splitting the spectrum into a high-capacity low-frequency encoder and a light high-frequency encoder.","keywords":["wind noise reduction","low-complexity neural network","dual-encoder architecture","real-time audio processing","embedded audio devices","single-channel audio enhancement","spectral characteristics of wind","mask-based noise suppression"],"falsifier":"Compute the spectral energy distribution of the recorded wind corpus: if the fraction of wind energy above 4 kHz is non-negligible (for example, more than a few percent), or if a held-out set of high-frequency-dominant wind recordings such as fan blast, propeller wash, or wind gust against a microphone grill pushes WindNetLite's leakage or SI-SDR clearly below WindNet's, the central assumption is falsified.","tokens_in":8080,"feed_emoji":"🎧","tokens_out":8385,"duration_ms":83784,"temperature":0.7,"pith_summary":"Wind noise is one of the most common outdoor recording problems, but suppressing it in real time is hard on the small processors inside phones and wearables. This paper tries to show that the task becomes cheap once you exploit a spectral fact: wind noise lives almost entirely in low frequencies, below about 4 kHz. The authors build WindNetLite, a 249K-parameter network that spends most of its capacity on a low-frequency encoder and only a little on the high-frequency sub-bands, and report that it matches the quality of the larger ULCNet and WindNet models while running at 73 MHz, about 40% of ULCNet's real-time cost. They also show that the choice of how the mask is used matters: estimating the wind and subtracting it distorts the target less, while directly rejecting wind suppresses it more.","feed_headline":"WindNetLite matches bigger noise reducers at 40% the compute","feed_subtitle":"A 249K-parameter dual-encoder network brings real-time wind-noise reduction to phones and wearables.","key_machinery":"The load-bearing mechanism is the frequency-split dual encoder. Channel-wise feature reorientation divides the power-law-compressed STFT magnitude into $N=10$ sub-bands; the first five, which are assumed to contain essentially all wind energy below about 4 kHz, feed a high-capacity convolutional encoder with a GRU, while the remaining sub-bands are average-pooled and fed to a much smaller encoder. This asymmetry is what buys the complexity reduction. Both streams are concatenated into an intermediate magnitude mask, which a lightweight two-layer CNN refines into a complex mask applied either to the desired signal (wind rejection) or to the wind component (wind extraction), with the power-law factor $\\alpha$ tuned separately for the two modes.","core_discovery":"WindNetLite is a causal, single-channel, frequency-split refinement of the ULCNet architecture. It maps the power-law-compressed STFT magnitude into ten sub-bands, sends the first five through a four-layer convolutional encoder followed by a temporal GRU, and sends the remaining sub-bands through a pooled three-layer encoder, then combines the two feature streams to produce an intermediate real mask that a lightweight CNN sharpens into a complex mask. With 249K parameters and roughly 73 MHz, it attains a real-time factor of 0.051 on an embedded-class single core, which is 40% of ULCNet's 0.127 and 84% of the 365K-parameter WindNet baseline's 0.061. On the speech test set it reaches PESQ 2.23 and SI-SDR 14.31 dB in extraction mode and leakage -2.81 in rejection mode; on instrumental music it reaches PEAQ -2.35 in rejection mode and SI-SDR 7.67 dB in extraction mode, close to WindNet on every metric. The paper also reports that rejection suppresses wind more strongly while extraction distorts the desired signal less, and that the optimal power-law compression factor is mode-dependent: 0.3 for rejection and 1.0 for extraction.","pith_inferences":["A direct spectral measurement of the 10-hour wind corpus is the natural check on the 4 kHz assumption; if high-frequency wind energy is non-negligible, the frequency split should be moved or made adaptive, an extension the paper leaves untested.","An ablation that gives the HF encoder the same capacity as the LF encoder would quantify how much of the quality parity comes from the frequency split itself rather than from overall model design.","The extraction/subtraction result suggests that for very low SNR wind, a two-stage pipeline that estimates wind first and then applies a post-filter, rather than direct masking, may generalize to other non-stationary noises."],"forward_implications":["WindNetLite runs at a real-time factor of 0.051 on an embedded-class single core, so causal wind-noise reduction fits within the compute budget of phones, wearables, action cameras, and hearables.","The extraction mode consistently yields higher SI-SDR, meaning less distortion to the desired audio, while the rejection mode yields lower leakage, meaning stronger wind suppression, giving system designers a direct knob to trade one objective against the other.","The same dual-encoder idea should transfer to other noise types with a known spectral footprint: put capacity in the sub-bands where the noise lives and keep the rest cheap.","The mode-dependent power-law factor means a single universal $\\alpha$ is not optimal for wind-noise reduction; tuning $\\alpha$ per reconstruction strategy is part of the performance gain.","Because the model is causal and single-channel, it can be inserted into existing audio capture chains without requiring multiple microphones or look-ahead buffering."],"supporting_citations":[{"why":"Supplies the ULCNet architecture, power-law STFT compression, and two-stage masking pipeline that WindNetLite modifies.","marker":"[13]"},{"why":"Supplies the AudioSet audio events mixed with recorded wind to build the 200-hour training and validation sets.","marker":"[21]"},{"why":"Supplies the VCTK speech utterances used as the desired signals in the speech test set.","marker":"[22]"},{"why":"Supplies the MUSDB18 instrumental music used as the desired signals in the music test set.","marker":"[23]"},{"why":"Defines the PEAQ metric used to score quality on the instrumental music test set.","marker":"[24]"},{"why":"Defines the PESQ metric used to score speech quality on the speech test set.","marker":"[25]"},{"why":"Defines the SI-SDR metric used to measure signal fidelity of the recovered target.","marker":"[26]"},{"why":"Provides the WebMUSHRA framework for the subjective listening test in Fig. 4.","marker":"[27]"}],"fun_headline_variants":["WindNetLite: 249K params matches ULCNet at 40% compute","Tiny 73-MHz wind reducer matches ULCNet at 40% compute","Low-complexity wind net: 40% compute, same quality as ULCNet","Real-time wind noise reduction on a 249K-param embedded net"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The architecture assumes wind noise is concentrated below roughly 4 kHz with negligible energy above it; if real wind or microphone setups produce substantial high-frequency wind energy, the small high-frequency encoder would miss it and the performance claim would weaken.","fun_headline_variants_meta":{"raw":{"variants":["WindNetLite: 249K params matches ULCNet at 40% compute","Tiny 73-MHz wind reducer matches ULCNet at 40% compute","Low-complexity wind net: 40% compute, same quality as ULCNet","Real-time wind noise reduction on a 249K-param embedded net"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001268,"raw_usage":{"total_tokens":5163,"prompt_tokens":889,"completion_tokens":4274,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":4184}},"tokens_in":505,"tokens_out":4274,"duration_ms":32220,"temperature":1.0,"reasoning_tokens":4184,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:41:35.113244+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the spectral energy distribution of the recorded wind corpus: if the fraction of wind energy above 4 kHz is non-negligible (for example, more than a few percent), or if a held-out set of high-frequency-dominant wind recordings such as fan blast, propeller wash, or wind gust against a microphone grill pushes WindNetLite's leakage or SI-SDR clearly below WindNet's, the central assumption is falsified.","supporting_citations":[{"cited_title":"Ultra low complexity deep learning based noise suppres- sion,","cited_arxiv_id":null,"evidence_quote":"Supplies the ULCNet architecture, power-law STFT compression, and two-stage masking pipeline that WindNetLite modifies."},{"cited_title":"Au- dio set: An ontology and human-labeled dataset for audio events,","cited_arxiv_id":null,"evidence_quote":"Supplies the AudioSet audio events mixed with recorded wind to build the 200-hour training and validation sets."},{"cited_title":"CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR V oice Cloning Toolkit (version 0.92), [sound].,","cited_arxiv_id":null,"evidence_quote":"Supplies the VCTK speech utterances used as the desired signals in the speech test set."},{"cited_title":"The MUSDB18 corpus for music separation,","cited_arxiv_id":null,"evidence_quote":"Supplies the MUSDB18 instrumental music used as the desired signals in the music test set."},{"cited_title":"PEAQ-The ITU standard for objective measurement of perceived audio quality,","cited_arxiv_id":null,"evidence_quote":"Defines the PEAQ metric used to score quality on the instrumental music test set."},{"cited_title":"Perceptual evaluation of speech quality (PESQ)- a new method for speech quality assessment of tele- phone networks and codecs,","cited_arxiv_id":null,"evidence_quote":"Defines the PESQ metric used to score speech quality on the speech test set."},{"cited_title":"SDR–half-baked or well done?,","cited_arxiv_id":null,"evidence_quote":"Defines the SI-SDR metric used to measure signal fidelity of the recovered target."},{"cited_title":"Towards the next generation of web-based experiments: A case study assessing basic audio quality following the itu-r recommen- dation bs. 1534 (mushra),","cited_arxiv_id":null,"evidence_quote":"Provides the WebMUSHRA framework for the subjective listening test in Fig. 4."}],"review_version":1}