{"id":"4e4fbbf2-8ec6-4cb1-99f1-3705dba4720a","arxiv_id":"2509.05079","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A 0.45M-parameter causal UNet-style denoiser with look-back frames and GRUs reports 22.34 dB SI-SDR on full-band VCTK data and RTF 0.014 on a Pixel 7.","lead":"This paper introduces a causal, low-latency DNN for full-band (48 kHz) speech denoising that runs on a mobile phone with a real-time factor below 0.02. Its headline result is a 22.34 dB SI-SDR on the VCTK test set, above reported PercepNet and DeepFilterNet scores, although PESQ-WB and DNSMOS are lower than one competitor.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline SI-SDR values are quoted under different test protocols, and Table I's arithmetic gives 5.71 dB, not the claimed 'over 6 dB'.","rationale":"The reader's weakest assumption is exactly the load-bearing concern I identify: baseline numbers are not computed under the same protocol, and the 6 dB claim rests on that comparison. In good faith, the paper is technically detailed and the architecture appears plausible, but the empirical headline is under-supported without a controlled baseline evaluation. I additionally note a concrete internal arithmetic problem: the only directly comparable SI-SDR numbers are 22.34 and 16.63, a difference of 5.71 dB, not 'over 6 dB'. This does not overturn the method's value, but it reinforces the need for verification. Since the reader already issued a CONDITIONAL verdict based on the same issue, my stress-test does not move the verdict; it stays CONDITIONAL, formally represented here as UNCHANGED. No code or checkpoints are released, but that is not an independent attack; it is part of why the baseline numbers cannot be audited. If the recomputation test is run and the gap holds, the verdict should move to ACCEPT.","tokens_in":9294,"tokens_out":4589,"duration_ms":54264,"concrete_test":"Run the official pretrained DeepFilterNet v1 (and v3, if a checkpoint is available) on the exact VCTK test-set mixtures used for Table I, using the same noise-mixing code, SNR draws, peak scaling, STFT parameters, and SI-SDR implementation as the proposed method. If the recomputed DFv1 SI-SDR changes by more than about 0.5 dB, or if the gap to the proposed method falls below 6 dB, then the 'over 6 dB' superiority claim is not supported by the current evidence. If the reproduced value matches 16.63, the protocol-mismatch concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section IV is that the method achieves the highest SI-SDR, with a difference of over 6 dB, while having the lowest complexity and latency. The only SI-SDR baseline shown is DFv1, whose 16.63 is taken from the DeepFilterNet paper, not recomputed under the same protocol used for the proposed method. The paper does not state that the baselines were evaluated with the same SNR distribution (−10 to 25 dB), the same peak scaling (0.001–0.999), the same noise-mixing procedure, or the same STFT/SI-SDR implementation. Section III-C only lists the comparison targets; Table I reports single numbers without variance or repeated trials. If the published DFv1 value was produced under a different SNR range, a different VCTK test split, or different evaluation code, the 5.7 dB gap could be partly a protocol artifact. Moreover, the text says 'over 6 dB', but the table shows 22.34 − 16.63 = 5.71 dB, which is not over 6 dB even with rounding. This arithmetic discrepancy makes the need for a direct recomputation more pressing. The cross-device RTF comparison is also fragile, but the baseline SI-SDR comparison is the most load-bearing premise for the headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a causal, full-band (48 kHz) speech denoising DNN intended for mobile deployment. The architecture is a U-Net variant with a learned input/output mapping, look-back frames at the encoder and mask predictor, temporally spanning convolutional kernels, and GRU-based bottlenecks, yielding 0.451M parameters, 0.0064 GMACs/s, and a measured TFLite RTF of 0.014 on a Pixel 7. Training uses SI-SDR as the loss on DNS-Challenge data filtered by DNSMOS, with VCTK as the test set. The paper reports SI-SDR of 22.34 dB, STOI 0.94, PESQ-WB 2.82, and claims a >6 dB SI-SDR advantage over DeepFilterNet v1 while being far lighter and faster. The authors also conjecture why PESQ-WB and DNSMOS are lower than DeepFilterNet v3.","tokens_in":9790,"tokens_out":1917,"duration_ms":24074,"significance":"If the headline claim survives scrutiny, the paper would be a genuinely useful engineering contribution: it demonstrates that a very small causal model can achieve SI-SDR competitive with or exceeding much larger full-band systems at 48 kHz, and it backs the deployability claim with an actual TFLite benchmark on a mobile phone. The explicit design choices--look-back frames for short temporal context, GRUs for long context, MobileNet-style bottlenecks, causal instance normalization--are clearly described and mostly reproducible from the text. The paper also uses public datasets and a conventional train/validation/test split with early stopping, which is good empirical hygiene. The main weakness is the comparison protocol: the baseline numbers are quoted from other papers rather than recomputed under the same test conditions, and the paper gives no variance or significance information for its own single-run results. This makes the central 'over 6 dB' claim fragile regardless of how carefully the proposed model is evaluated.","major_comments":[{"comment":"The SI-SDR baseline for DeepFilterNet v1 (16.63 dB) is taken from the DeepFilterNet paper, not recomputed under the protocol used for the proposed method. The training/test protocol in Section III-A differs from other works in SNR range (-10 to 25 dB), peak scaling (0.001-0.999), noise mixing (up to two noise files), and STFT parameters. Since SI-SDR is sensitive to these choices, the 5.71 dB gap shown in Table I could be partly a protocol artifact. The authors should either recompute all baselines under the identical evaluation script/protocol or clearly label the comparison as cross-paper and temper the claim accordingly.","section":"Section III-C and Table I"},{"comment":"The text states 'our method provides the highest SI-SDR with a difference of over 6 dB.' Table I gives Ours = 22.34 and DFv1 = 16.63, so the difference is 5.71 dB, not over 6 dB. This arithmetic discrepancy matters because the >6 dB gap is the headline quantitative claim. It should be corrected and the claim restated based on the actual recomputed or cross-paper numbers.","section":"Section IV, 'over 6 dB' claim"},{"comment":"The paper reports a single run for the proposed method and no confidence intervals, error bars, or significance tests for any metric. Given that the central claim is a numerical advantage over baselines, the absence of any measure of variability (e.g., multiple training runs, bootstrap intervals, or paired significance tests) weakens the support. At minimum, the authors should state whether the result is from one run, report the validation performance, and add variance estimates or statistical testing.","section":"Section III-C and Table I (single-run, no uncertainty)"},{"comment":"The RTF comparison is across different hardware: Ours is measured on a Pixel 7 phone, while PercepNet and DFv1 values are taken from a laptop-based reference. The Table II note acknowledges this, but the conclusion 'almost as four times faster' is not supported by a like-for-like comparison. The authors should either benchmark the baselines on the same mobile device or explicitly delimit the claim to 'RTF on different hardware' and avoid direct speed superiority statements.","section":"Section III-C and Table II"}],"minor_comments":[{"comment":"The explanation for lower PESQ-WB (that PESQ only evaluates up to 8 kHz and penalizes fullband artifacts) is presented as a likely cause but is speculative. If the authors want to make this claim, they should provide supporting evidence, e.g., band-limited analysis or an ablation.","section":"Section IV"},{"comment":"The DNSMOS filtering criterion is given, but the DNSMOS model version and the exact threshold rationale are not discussed. Also, the paper does not state how many files were removed by filtering.","section":"Section III-A"},{"comment":"The phrase 'having 16 Hz sampling rate' in the Introduction appears to be a typo for 16 kHz. Similarly, 'denoinsing' and 'wtih' typos appear in the text.","section":"Section III-B"},{"comment":"The comparison with PercepNet is limited because PercepNet is not evaluated on SI-SDR or STOI in Table I; only PESQ-WB is listed. This makes the 'surpasses all but one' claim only partially supported.","section":"Section III-C"},{"comment":"The notation for the dimensions of H_in-pad would benefit from a concrete tensor-shape example. Also, the exact implementation of the 'causal instance norm' (whether it is a running-statistics or batch-dependent normalization) is not fully specified.","section":"Section II"}],"recommendation":"major_revision","confidential_remarks":"The paper's engineering message is plausible and useful, but its headline claim rests on cross-paper comparisons that are not apples-to-apples. I think the authors can address this within revision by recomputing baselines under their own protocol or by clearly marking the comparison as indicative and correcting the 'over 6 dB' statement. The single-run issue also needs to be acknowledged and ideally quantified. I would not reject, because the proposed system's complexity and latency figures are concrete and the architecture is original enough for a venue focused on practical speech enhancement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The takeaway: this is a useful engineering paper, not a scientific breakthrough. The authors build a 0.45M-parameter causal full-band (48 kHz) denoiser that runs on a phone with RTF 0.014 and reports 22.34 dB SI-SDR on VCTK. If those numbers survive reproduction, the architecture is a practical addition to the mobile voice stack. The design is a sensible combination of known pieces—UNet with MobileNet inverted bottlenecks, look-back frames, a GRU bottleneck, and a GRU-based mask autoencoder. That combination is new in this exact form, and the paper describes it clearly enough to reimplement.\n\nThe evaluation is mostly sound: public datasets, an 80/20 split, early stopping, and the authors are careful to exclude VCTK overlap between training and test. The DNSMOS-based filtering of training data is a reasonable pre-processing choice. They also explicitly note that the RTF comparison across devices might be pessimistic for their method, which is honest.\n\nThe soft spots are real but fixable. The headline claim—'highest SI-SDR, over 6 dB'—rests on comparing against DFv1's published value of 16.63, which was likely computed under a different test protocol (different SNR distribution, peak scaling, or evaluation code). The paper does not rerun the baselines, reports no variance, and does no significance testing. The arithmetic is also off: 22.34 - 16.63 = 5.71 dB, not over 6 dB. That matters because the text makes the claim explicitly. Also, their PESQ-WB and DNSMOS are lower than DeepFilterNet v3; they conjecture this is because those metrics are 16 kHz-centric, which is plausible but not demonstrated. That's an honest limitation, not a deception.\n\nWho is this for? Practitioners in real-time full-band enhancement on constrained hardware. A serious referee should handle it; the core feasibility claim is credible, and the engineering detail is worth scrutiny. The authors should be asked to recompute or clearly state the baseline protocol, report standard deviations, fix the arithmetic, and ideally release a checkpoint.","headline":"Solid engineering result with a credible mobile full-band denoiser, but the headline SOTA claim rests on cross-protocol baseline numbers and an arithmetic slip (5.71 dB, not 'over 6').","tokens_in":10209,"tokens_out":2300,"would_cite":false,"duration_ms":24445,"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 causal 48 kHz denoiser with 0.45M parameters is reported to beat full-band baselines by over 6 dB SI-SDR while running at real-time factor 0.014 on a phone CPU.","keywords":["speech denoising","full-band audio","low latency","real-time","UNet","recurrent neural networks","mobile devices","SI-SDR"],"falsifier":"Run the three baselines through the exact test pipeline used here (VCTK test set, SNR sampled uniformly from -10 to 25 dB, up to two noise files mixed, peak amplitude scaled to 0.001-0.999, 2048-point STFT, and the same SI-SDR implementation). If the proposed model's SI-SDR does not exceed the baselines by roughly 6 dB, or falls below them, the central claim is refuted.","tokens_in":9255,"feed_emoji":"🎧","tokens_out":14945,"duration_ms":124401,"temperature":0.7,"pith_summary":"The paper sets out to show that full-band (48 kHz) speech denoising can be done in real time on a mobile phone without giving up quality. It builds a causal, low-latency DNN that processes one short-time Fourier transform (STFT) magnitude frame at a time, with no look-ahead frames, and reports the highest scale-invariant signal-to-distortion ratio (SI-SDR) among full-band methods it compares against: 22.34 dB on the VCTK test set, more than 6 dB above published results for PercepNet, DeepFilterNet v1, and DeepFilterNet v3. That result comes with 0.45M parameters, 0.0064 GMACs, and a real-time factor of 0.014 on a phone CPU, roughly 100x fewer MACs and about 4x lower latency than the compared baselines. If the comparison holds, this closes the usual gap between high-quality full-band denoising and deployment on resource-constrained devices.","feed_headline":"A causal 48 kHz denoiser beats full-band baselines by over 6 dB SI-SDR","feed_subtitle":"Real-time factor 0.014 on a phone CPU and no look-ahead frames, so it can slot into full-band voice chains.","key_machinery":"The load-bearing mechanism is a pair of temporal-context techniques used twice. In the front end, 32 look-back frames of the mapped STFT magnitude are stacked along time and folded into the channel dimension by temporally-spanning convolutional kernels, capturing short patterns; in the bottleneck and again just before mask prediction, single-layer GRUs carry state across frames, capturing long patterns. The resulting causal system never sees a future frame. Complexity is kept low by a learned input mapping that shrinks the full STFT to 96 features, MobileNet-v2-style inverted bottlenecks in the encoder, and mask prediction in that reduced space, with the mask then mapped back to all frequenc","core_discovery":"The central claim is that a deliberately minimal, causal full-band denoiser can exceed the SI-SDR of established full-band methods while being much smaller and faster. The model takes the current STFT magnitude frame plus 32 prior mapped frames, encodes them with a UNet-style stack whose kernels span the look-back window, runs a GRU bottleneck to carry long-term state, decodes with skip connections, then passes the representation through a second small GRU autoencoder before predicting a sigmoid mask in a reduced 96-dimensional space. The mask multiplies the noisy magnitude; phase passes through unchanged. On the VCTK test set the authors report 22.34 dB SI-SDR, 0.94 STOI, and 2.82 PESQ-WB,","pith_inferences":["If the 6 dB SI-SDR gap survives a shared evaluation protocol, temporal modeling (look-back plus two GRU stages) is likely the main driver of the gain, meaning similar advantages might transfer to other full-band tasks such as dereverberation or acoustic echo cancellation.","A direct test of the metric hypothesis would be to recompute SI-SDR after low-pass filtering both the estimate and reference to 8 kHz; if the gap mostly remains, the PESQ-WB discrepancy is not simply a bandwidth artifact.","Because phase is unmodified and only magnitude is masked, the method can be dropped into an existing codec or enhancement pipeline; measuring end-to-end quality, rather than isolated denoising quality, would show whether the low perceptual scores matter in practice.","The learned input mapping to 96 features buys speech fidelity at a roughly 14% speed cost; replacing it with a fixed filterbank in an ablation would separate the benefit of learned compression from the benefit of the temporal architecture."],"forward_implications":["A causal, no-look-ahead model can deliver SI-SDR gains of more than 6 dB over full-band baselines, so real-time voice chains no longer need to sacrifice denoising quality to stay causal.","At 0.45M parameters and 0.0064 GMACs, full-band 48 kHz denoising fits in mobile budgets and can run on a phone CPU at a real-time factor of 0.014, leaving headroom for the rest of the processing chain.","Removing the learned input/output mapping cuts the real-time factor to about 0.012, showing that the mapping itself contributes roughly 14% of the runtime cost.","The reported gap between high SI-SDR and lower PESQ-WB suggests that full-band methods need metrics that reward the 8-24 kHz region, not just narrowband quality."],"supporting_citations":[{"why":"PercepNet, the full-band low-complexity baseline whose published PESQ-WB (2.54) and design are the comparison targets for the headline SI-SDR advantage.","marker":"[11]"},{"why":"DeepFilterNet v1, the full-band deep-filtering baseline that supplies the SI-SDR 16.63, STOI 0.94, PESQ-WB 2.81, and the 1.78M/0.35G/RTF 0.04 complexity figures the paper beats.","marker":"[2]"},{"why":"DeepFilterNet v3, the perceptual-score leading baseline (PESQ-WB 3.17, MSIG 4.34, MOVRL 3.77) whose scores frame the paper's lower PESQ-WB discussion.","marker":"[23]"},{"why":"DNS-Challenge dataset used as the training/validation source, from which clean speech and noise are resampled to 48 kHz and mixed.","marker":"[13]"},{"why":"VCTK test set, the 48 kHz evaluation data on which the reported SI-SDR, STOI, PESQ-WB, and DNSMOS values are computed.","marker":"[15]"},{"why":"DNSMOS non-intrusive speech-quality metric used both to filter the dev-set (MOS-SIG<4.0, MOS-OVRL<3.9) and to report MSIG/MOVRL results.","marker":"[16]"},{"why":"Defines scale-invariant SDR, the headline metric and the training loss used to optimize the model.","marker":"[20]"},{"why":"DeepFilterNet2 paper, the source of the PercepNet and DeepFilterNet v1 real-time factors used in Table II's complexity comparison.","marker":"[24]"},{"why":"MobileNetV2 inverted bottlenecks, the complexity-reduction building blocks of the encoder.","marker":"[12]"},{"why":"U-Net architecture with skip connections on which the encoder-decoder is based.","marker":"[9]"}],"fun_headline_variants":["Causal 48 kHz denoiser: 22.34 dB SI-SDR, runs on phones","Lightweight UNet+GRU hits 22.34 dB SI-SDR at 48 kHz","Full-band denoiser: UNet+GRU, 22.34 dB, runs real-time on phones","48 kHz denoiser: causal, low-latency, beats bigger models"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The headline 6 dB SI-SDR advantage assumes the PercepNet, DeepFilterNet v1, and DeepFilterNet v3 numbers quoted from other papers came from the same test protocol—same SNR sampling, noise mixing, peak scaling, and SI-SDR code—as the proposed method; if the published baselines were evaluated under different conditions, the reported gap could be an artifact of protocol mismatch rather than model superiority.","fun_headline_variants_meta":{"raw":{"variants":["Causal 48 kHz denoiser: 22.34 dB SI-SDR, runs on phones","Lightweight UNet+GRU hits 22.34 dB SI-SDR at 48 kHz","Full-band denoiser: UNet+GRU, 22.34 dB, runs real-time on phones","48 kHz denoiser: causal, low-latency, beats bigger models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000653,"raw_usage":{"total_tokens":2860,"prompt_tokens":808,"completion_tokens":2052,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":1949}},"tokens_in":552,"tokens_out":2052,"duration_ms":15349,"temperature":1.0,"reasoning_tokens":1949,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:36:00.043593+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the three baselines through the exact test pipeline used here (VCTK test set, SNR sampled uniformly from -10 to 25 dB, up to two noise files mixed, peak amplitude scaled to 0.001-0.999, 2048-point STFT, and the same SI-SDR implementation). If the proposed model's SI-SDR does not exceed the baselines by roughly 6 dB, or falls below them, the central claim is refuted.","supporting_citations":[{"cited_title":"A perceptually-motivated approach for low-complexity, real-time enhancement of fullband speech,","cited_arxiv_id":null,"evidence_quote":"PercepNet, the full-band low-complexity baseline whose published PESQ-WB (2.54) and design are the comparison targets for the headline SI-SDR advantage."},{"cited_title":"Deepfil- ternet: A low complexity speech enhancement framework for full-band audio based on deep filtering,","cited_arxiv_id":null,"evidence_quote":"DeepFilterNet v1, the full-band deep-filtering baseline that supplies the SI-SDR 16.63, STOI 0.94, PESQ-WB 2.81, and the 1.78M/0.35G/RTF 0.04 complexity figures the paper beats."},{"cited_title":"Deepfilternet: Perceptually motivated real-time speech enhancement,","cited_arxiv_id":null,"evidence_quote":"DeepFilterNet v3, the perceptual-score leading baseline (PESQ-WB 3.17, MSIG 4.34, MOVRL 3.77) whose scores frame the paper's lower PESQ-WB discussion."},{"cited_title":"ICASSP 2023 Deep Noise Suppression Challenge,","cited_arxiv_id":null,"evidence_quote":"DNS-Challenge dataset used as the training/validation source, from which clean speech and noise are resampled to 48 kHz and mixed."},{"cited_title":"Investi- gating rnn-based speech enhancement methods for noise-robust text-to- speech,","cited_arxiv_id":null,"evidence_quote":"VCTK test set, the 48 kHz evaluation data on which the reported SI-SDR, STOI, PESQ-WB, and DNSMOS values are computed."},{"cited_title":"DNSMOS: A Non-Intrusive Perceptual Objective Speech Quality metric to evaluate Noise Suppres- sors,","cited_arxiv_id":null,"evidence_quote":"DNSMOS non-intrusive speech-quality metric used both to filter the dev-set (MOS-SIG<4.0, MOS-OVRL<3.9) and to report MSIG/MOVRL results."},{"cited_title":"SDR – Half- baked or Well Done?","cited_arxiv_id":null,"evidence_quote":"Defines scale-invariant SDR, the headline metric and the training loss used to optimize the model."},{"cited_title":"Deepfil- ternet2: Towards Real-Time Speech Enhancement on Embedded Devices for Full-Band Audio,","cited_arxiv_id":null,"evidence_quote":"DeepFilterNet2 paper, the source of the PercepNet and DeepFilterNet v1 real-time factors used in Table II's complexity comparison."},{"cited_title":"MobileNetV2: Inverted Residuals and Linear Bot- tlenecks,","cited_arxiv_id":null,"evidence_quote":"MobileNetV2 inverted bottlenecks, the complexity-reduction building blocks of the encoder."},{"cited_title":"U-Net: Convolutional Net- works for Biomedical Image Segmentation,","cited_arxiv_id":null,"evidence_quote":"U-Net architecture with skip connections on which the encoder-decoder is based."}],"review_version":1}