{"id":"ad200c1e-7604-4431-b287-96b98e9a97a3","arxiv_id":"2505.23515","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A 3.58M-parameter, two-stage causal speech enhancer adds a GAN second stage to DeepFilterNet2 and improves NISQA-MOS on the URGENT test set.","lead":"This paper combines a standard predictive speech denoiser with a small GAN-based second stage to restore speech details the first stage removes. It reports better quality scores on a public test set with only 3.58M parameters and 40 ms latency, targeting real-time streaming use.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported NISQA gain may be an artifact of comparing against a weakened retrained first stage, not of the proposed two-stage framework.","rationale":"The reader's weakest_assumption focuses on the mechanism of noisy conditioning and whether the GAN learns P(x|y,z) rather than merely shifting the distribution. That is a legitimate concern, but the most load-bearing worry is about the baseline: the only head-to-head comparison supporting the central improvement claim is against a retrained first stage that is 0.62 NISQA-MOS worse than the published DeepFilterNet2. The paper's system is DeepFilterNet2 plus a GAN, and its own Table 1 shows that the proposed full system (3.12) does not beat the published DeepFilterNet2 (3.28) on NISQA-MOS; it only beats the retrained and weaker first stage. Therefore the central claim as stated, 'our system improves over the first stage in terms of NISQA-MOS', is true for the retrained stage but the meaningful comparison against the strong published first stage is absent. This is a correctness risk, not a novelty judgment: the contribution's value depends on the second stage adding genuine gain over a strong predictive backbone. The proposed concrete test settles it directly. I agree partially with the reader's mechanism concern, but I would prioritize the baseline comparison because the ablation without noisy concat (2.86) is still above the retrained first stage (2.66), which actually supports the mechanism somewhat; the baseline gap is the place where the argument is least secure.","tokens_in":7726,"tokens_out":1638,"duration_ms":13148,"concrete_test":"Retrain or reuse the published DeepFilterNet2 checkpoint as the frozen first stage and train the identical second-stage GAN on top of it, then evaluate on the same 2024 URGENT non-blind test set. If the full system's NISQA-MOS minus the published DeepFilterNet2's NISQA-MOS is no longer positive (e.g., within noise of 3.28), the headline claim that the proposed system improves over the first stage is not established.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is a NISQA-MOS gain over the first stage: proposed DeepFilterGAN 3.12 vs. First Stage only 2.66 (Table 1). However, the baseline 'First Stage only (retrained [11])' is 0.62 NISQA-MOS below the published DeepFilterNet2 that the first stage is explicitly 'adopted from' (3.28). The paper never explains why its retrained first stage is so much weaker than the published checkpoint, and no code, hyperparameters, or training details beyond epochs (45) are given to permit replication. If the retrained first stage is suboptimally trained or under-fit, then the apparent improvement of the GAN second stage may be at least partly a compensation for a degraded first stage rather than a demonstration of stochastic regeneration. The claim in the abstract that 'our system improves over the first stage in terms of NISQA-MOS metric' would then hold trivially, but the more general claim that a lightweight GAN stage can improve a strong predictive enhancer would be unsupported. The published DeepFilterNet2 checkpoint is available, so this is directly testable: run the identical second-stage training on top of the published DeepFilterNet2. A second specific weakness is statistical: all NISQA-MOS numbers in Table 1 are single-point estimates with no confidence intervals, while NISQA is a non-intrusive model with known prediction variance.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DeepFilterGAN, a two-stage causal speech enhancement system. The first stage is a retrained DeepFilterNet2; the second stage is a GAN whose generator is based on OnlineSpatialNet and whose discriminator is based on MelGAN. The generator receives the concatenation of the noisy and intermediate-enhanced STFTs, which the authors call noisy conditioning. The system has 3.58M parameters and 40 ms algorithmic latency. Experiments on the 2024 URGENT non-blind test set report NISQA-MOS of 3.12 for the proposed system versus 2.66 for the retrained first stage and 2.86 for an ablation without noisy conditioning. The paper also compares with published DeepFilterNet2 (3.28) and UNIVERSE++ (3.44) and reports a favorable overall ranking.","tokens_in":7983,"tokens_out":8989,"duration_ms":75892,"significance":"If the central result holds, the paper demonstrates that a compact GAN second stage can improve a predictive enhancer's perceived quality without requiring multiple diffusion inference steps, which is useful for real-time and embedded speech enhancement. The evaluation uses an external challenge test set with multiple objective metrics, and the ablation isolates the contribution of noisy conditioning. Strengths include the low parameter count, the explicit algorithmic-latency statement, and the use of a public challenge protocol. However, the main empirical claim is weakened by the unexplained gap between the retrained first stage and the published DeepFilterNet2 checkpoint, and by missing statistical and reproducibility details.","major_comments":[{"comment":"The baseline used to demonstrate the central improvement, 'First Stage only (retrained [11])', obtains NISQA-MOS 2.66, which is 0.62 below the publicly released DeepFilterNet2 checkpoint (3.28) evaluated on the same test set. Since the first stage is explicitly adopted from DeepFilterNet2, the paper should explain why retraining on the 2025 URGENT training set yields such a large drop; the training description in Section 3.2 reports only '45 epochs' and omits the loss weights and optimizer details. Because the second stage is trained on top of the retrained first stage, the reported gain (2.66 to 3.12) may be partly compensating for an underfit first stage. The published DeepFilterNet2 checkpoint is available, so the authors can directly test this by training the identical second stage on the published first stage and reporting the resulting NISQA-MOS; alternatively, they should show that the retrained first stage is not anomalously weak.","section":"Table 1, Section 4"},{"comment":"The second-stage generator is described only as 'adopted from OnlineSpatialNet' with two input channels, four blocks, and hidden size 16. The paper does not state whether the generator outputs a complex mask, a complex spectrogram, or a magnitude estimate; it also omits the value of beta in Eq. (3), the learning rates for generator and discriminator, the gradient-clipping threshold, and the exact weight-normalization and update schedules beyond 'every second iteration'. Without these details, the experiments are not reproducible and the contribution of the specific generator design to the NISQA gain cannot be isolated. Please provide an architecture summary (layer types, kernel sizes, skip connections, Mamba configuration) and a complete hyperparameter table.","section":"Sections 3.1 and 3.2, Eq. (3)"},{"comment":"All metrics in Table 1 are single-point estimates with no confidence intervals or significance tests. The central claim relies on a NISQA-MOS increase of 0.46 over the first stage and an ablation difference of 0.26 for noisy concatenation, but NISQA is a non-intrusive predictor with nontrivial prediction variance, and several objective metrics in the same table move in the opposite direction (e.g., PESQ 2.07 to 2.03, word accuracy 75.51 to 75.05). At minimum, the authors should report bootstrap confidence intervals or per-utterance distributions for the principal metrics (especially NISQA-MOS, PESQ, and word accuracy) to establish that the reported improvements and non-degradations are not within the noise.","section":"Table 1 and Section 4"},{"comment":"The title and abstract describe the system as real-time, but the paper reports only the 40 ms algorithmic latency from the two-frame lookahead; no real-time factor (RTF) or measured inference latency on any hardware is provided. For a streaming/embedded system, the RTF relative to real time on the target platform is a standard requirement and is load-bearing for the 'real-time' claim. Please report the RTF or wall-clock latency for at least one representative device (e.g., a notebook CPU or an embedded ARM core).","section":"Section 3.2 and title"}],"minor_comments":[{"comment":"The text states that each MelGAN discriminator 'focuses on a different frequency range', but the original MelGAN multi-scale discriminators operate on different time scales (via sub-sampling of the waveform), not frequency ranges. Please correct the description.","section":"Section 2.4"},{"comment":"The sentence 'We use the same learning rate and weight decay schedule as in [11]' is ambiguous: it appears after the description of second-stage training, but it is unclear whether it applies to the first-stage training (which is described only by '45 epochs') or only to the GAN training. Please state explicitly which learning-rate and weight-decay schedules are used for each training step.","section":"Section 3.2"},{"comment":"The row label 'First Stage only (retrained [11])' is confusing because the published DeepFilterNet2 row is also present. Rename it to something like 'First stage only (retrained by authors)' to make the distinction explicit.","section":"Table 1"},{"comment":"The overall ranking in Eq. (4) is computed over only five systems, so the difference between ranking scores (e.g., 2.25 vs 2.31) is coarse; the text should note that the ranking interpretation is limited by the small number of compared systems.","section":"Section 4"},{"comment":"The green-boxed region in Figure 3 is described only by color; if the manuscript is printed in grayscale, the comparison is hard to follow. Add a text label or a distinct line style to mark the recovered segment.","section":"Section 4, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a conference-style challenge paper (six pages with index terms) submitted to a journal. The main technical concern is the unexplained gap between the retrained and published first-stage checkpoints; if the authors can close that gap or demonstrate the gain with the published checkpoint, the paper would be much more convincing. There is no evidence of fabrication; the evaluation uses a public test set. The lack of code and hyperparameters is a concern for reproducibility but should be fixable in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Overall: this is a credible incremental engineering paper. The genuinely new piece is the noisy-conditioning trick—feeding both the noisy input and the first-stage output as a two-channel input to a Mamba-based generator—and the ablation showing it buys about 0.26 NISQA-MOS (2.86 to 3.12). That ablation is clean, and the writing is honest about the frozen-first-stage limitation. The system is small (3.58M parameters) and the 40ms algorithmic latency is plausible for streaming.\n\nThe main soft spot is the baseline. The paper compares against a retrained first stage that scores 2.66 NISQA-MOS on the URGENT 2024 test set, while the published DeepFilterNet2 checkpoint scores 3.28. The retrained model is not the same as the published model—45 epochs on URGENT data—so the 0.46 MOS gain from the GAN stage may partly be compensating for an under-trained first stage rather than demonstrating that a GAN can improve a strong predictive enhancer. The abstract's claim \"improves over the first stage\" is literally true, but the more general claim is unsupported. This is directly testable: run the same second-stage training atop the published checkpoint. Until that is done, the result should be read as \"a GAN can patch a weak predictive model,\" not \"GAN stochastic regeneration strengthens a strong predictive model.\"\n\nOther soft spots are minor but real: no confidence intervals or significance tests on the NISQA scores, no actual runtime measurements despite the real-time claim, and no code or audio. The loss weighting beta in Eq. 3 is not reported, so the training is not fully reproducible. The citation pattern is fine; the related work is appropriate.\n\nThe paper is not a breakthrough, but it is a solid engineering data point. It would be of value to people building lightweight real-time enhancement for embedded devices, and to anyone curious about whether GANs can substitute for diffusion inside stochastic regeneration. It deserves a serious referee, but the missing comparison and significance testing should be requested before acceptance. If the authors run the second stage on the published DeepFilterNet2 and report variance, this becomes a clean contribution. I would not cite it as-is for a strong claim, but I would accept it for review.","headline":"A competent engineering paper with a clean noisy-conditioning ablation, but the headline MOS gain is measured against a weakened retrained first stage, so the broader claim about improving a strong predictive enhancer is not yet supported.","tokens_in":8550,"tokens_out":3095,"would_cite":false,"duration_ms":29194,"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":"DeepFilterGAN claims that a full-band, real-time speech enhancer can improve perceived quality by regenerating speech from a two-channel input of noisy and intermediate-enhanced spectrograms.","keywords":["speech enhancement","stochastic regeneration","generative adversarial network","real-time processing","full-band audio","deep filtering","noisy conditioning","URGENT challenge"],"falsifier":"Run the two variants on test utterances where the first stage removes entire phonemes; if the no-concat variant still recovers those phonemes, or if replacing the noisy channel with a fixed constant keeps NISQA-MOS near 3.12, the paper's noisy-conditioning mechanism is not what carries the gain.","tokens_in":7488,"feed_emoji":"🎙️","tokens_out":11739,"duration_ms":96752,"temperature":0.7,"pith_summary":"DeepFilterGAN is a full-band speech enhancement system that pairs a predictive first stage with a generative adversarial second stage. The paper's central claim is that this second stage, trained with the first stage frozen, improves perceived speech quality by regenerating the clean speech distribution from a two-channel input of the noisy signal and the first stage's output. On the 2024 URGENT non-blind test set, the system raises NISQA-MOS from 2.66 (first stage alone) to 3.12, with 3.58M parameters and 40 ms algorithmic latency. An ablation that removes the noisy input drops to 2.86, which the authors present as evidence that noisy conditioning is what enables the gain. This matters because generative restoration has usually required large models or many inference steps; here it fits inside a streaming, full-band system.","feed_headline":"GAN second stage lifts real-time speech quality from 2.66 to 3.12","feed_subtitle":"A 3.58M-parameter system improves perceived quality by conditioning a GAN on noisy and denoised spectrograms.","key_machinery":"Stochastic regeneration with two-channel conditioning. The first stage produces an intermediate enhanced spectrogram $Z(k,f)$; the generator receives $[Y(k,f), Z(k,f)]$ as a two-channel input and is trained to sample the clean speech distribution $P(x|y,z)$ under a multi-scale adversarial loss. The generator is a lightweight modification of Online SpatialNet (1.14M parameters) and the discriminator follows MelGAN's three-scale design; the final objective is $\\min_G \\mathbb{E}_{y,z}[\\sum_l -D_l(\\hat{x})] + \\beta\\|x-\\hat{x}\\|_1$, where $\\hat{x} = \\operatorname{istft}(G(y,z))$.","core_discovery":"The discovery is that stochastic regeneration with a GAN can repair over-suppression in a real-time predictive enhancer. Specifically, the generator receives the concatenated spectrograms $[Y(k,f), Z(k,f)]$ and is trained against three MelGAN-style discriminators with a hinge loss plus a time-domain L1 term, learning the conditional distribution $P(x|y,z)$. On the challenge's non-blind test set, the final system scores 3.12 NISQA-MOS versus 2.66 for the retrained first stage, holds PESQ, ESTOI, SDR, and LSD roughly constant, and obtains the best overall ranking among compared models. A spectrogram example shows a speech segment removed by the first stage partially reappearing only when the noisy input is included.","pith_inferences":["The noisy conditioning channel may act as more than noise context: it could anchor the generator to the input signal, and a scrambling experiment could separate those two roles.","Jointly training both stages, which the paper lists as future work, is a natural next step because a frozen first stage may limit how well the generator can learn $P(x|y,z)$; end-to-end training could widen the NISQA-MOS gap.","Because the second stage has only 1.14M parameters, the same noisy-concat conditioning could be tested on even smaller on-device models.","The time-domain L1 term in the generator objective probably explains why intrusive metrics stay stable; varying $\\beta$ would directly probe the trade-off between adversarial sharpness and signal conservation."],"forward_implications":["The full system improves NISQA-MOS from 2.66 to 3.12 over the retrained first stage on the 2024 URGENT non-blind test set, while other objective metrics stay roughly level.","Concatenating the noisy input with the first-stage output is load-bearing for the improvement: removing it lowers NISQA-MOS to 2.86 and severely reduces recovery of over-suppressed speech content.","A generative second stage can fit inside a real-time system: 3.58M training parameters (3.45M at inference) and 40 ms algorithmic latency.","Despite lower NISQA-MOS than a larger diffusion-based comparison system, DeepFilterGAN obtains the best overall ranking across nonintrusive, intrusive, downstream-task-independent, and downstream-task-dependent metric categories.","The system preserves intelligibility-side metrics (phoneme similarity and word accuracy rate) better than the larger diffusion-based comparison."],"supporting_citations":[{"why":"Supplies the predictive first-stage model (DeepFilterNet2), its architecture, and the training hyperparameter schedule reused in the paper.","marker":"[11]"},{"why":"Defines the stochastic regeneration framework that motivates combining a predictive stage with a generative stage.","marker":"[5]"},{"why":"Provides the generator backbone (Online SpatialNet) modified to accept the two-channel noisy-plus-enhanced input.","marker":"[12]"},{"why":"Supplies the multi-scale discriminator design and hinge loss used for the second stage's adversarial training.","marker":"[13]"},{"why":"Serves as the main comparison baseline, a larger universal enhancement system that the compact model is measured against.","marker":"[6]"},{"why":"Defines the 2024 URGENT non-blind test set and metric categories from which Table 1 is drawn.","marker":"[15]"},{"why":"Provides the 2025 URGENT training dataset covering seven distortion types used to train the system.","marker":"[16]"},{"why":"Defines the NISQA-MOS metric used as the paper's main improvement criterion.","marker":"[19]"}],"fun_headline_variants":["GAN repair lifts speech quality in real-time streaming","GAN's stochastic regeneration restores suppressed speech","Real-time GAN adds back speech removed by denoiser","GAN regenerates lost speech, lifts MOS to 3.12","Over-suppression repaired by GAN, real-time and lightweight"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a 1.14M-parameter generator can learn the clean-speech distribution from a two-channel input of noisy and first-stage-enhanced spectrograms, so that the measured NISQA-MOS gain is genuine restoration rather than a generic reshaping of the output.","fun_headline_variants_meta":{"raw":{"variants":["GAN repair lifts speech quality in real-time streaming","GAN's stochastic regeneration restores suppressed speech","Real-time GAN adds back speech removed by denoiser","GAN regenerates lost speech, lifts MOS to 3.12","Over-suppression repaired by GAN, real-time and lightweight"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000871,"raw_usage":{"total_tokens":3736,"prompt_tokens":877,"completion_tokens":2859,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":2779}},"tokens_in":493,"tokens_out":2859,"duration_ms":18786,"temperature":1.0,"reasoning_tokens":2779,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:44:02.138368+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two variants on test utterances where the first stage removes entire phonemes; if the no-concat variant still recovers those phonemes, or if replacing the noisy channel with a fixed constant keeps NISQA-MOS near 3.12, the paper's noisy-conditioning mechanism is not what carries the gain.","supporting_citations":[{"cited_title":"DeepFilterNet2: Towards real-time speech enhancement on em- bedded devices for full-band audio,","cited_arxiv_id":null,"evidence_quote":"Supplies the predictive first-stage model (DeepFilterNet2), its architecture, and the training hyperparameter schedule reused in the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the stochastic regeneration framework that motivates combining a predictive stage with a generative stage."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the generator backbone (Online SpatialNet) modified to accept the two-channel noisy-plus-enhanced input."},{"cited_title":"The discriminator learns to distinguish real clean speech from enhanced speech, while the generator refines its output to resemble clean speech more closely","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-scale discriminator design and hinge loss used for the second stage's adversarial training."},{"cited_title":"As a result, the generator learns the clean speech distribu- tion conditioned on the noisy speech and the intermediate en- hanced speech, i.e","cited_arxiv_id":null,"evidence_quote":"Serves as the main comparison baseline, a larger universal enhancement system that the compact model is measured against."},{"cited_title":"D2former: A fully complex dual-path dual- decoder conformer network using joint complex masking and complex spectral mapping for monaural speech enhancement,","cited_arxiv_id":null,"evidence_quote":"Provides the 2025 URGENT training dataset covering seven distortion types used to train the system."},{"cited_title":"Storm: A diffusion-based stochastic regeneration model for speech en- hancement and dereverberation,","cited_arxiv_id":null,"evidence_quote":"Defines the NISQA-MOS metric used as the paper's main improvement criterion."}],"review_version":1}