{"id":"4e64d679-d75c-4de7-a473-a46653d14319","arxiv_id":"1908.08044","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A coarse-to-fine cosine similarity loss with a dynamic perceptual loss improves objective speech enhancement scores on the Valentini benchmark.","lead":"This paper proposes a training method for speech enhancement that optimizes a cosine similarity loss at multiple audio scales, from whole utterances down to short segments, and extends the same idea to the adversarial loss in a conditional GAN. The reported gains on standard objective speech quality metrics are consistent but modest, making this a useful engineering contribution rather than a conceptual breakthrough.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The coarse-to-fine mechanism is not isolated from a plain multi-scale loss: no control runs all granularities at every epoch, and dynamic perceptual loss has no static feature-matching control.","rationale":"The reader's weakest_assumption identifies exactly the missing multi-scale control for the temporal schedule. I agree and add that the same missing control applies to dynamic perceptual loss. The paper's central contribution is the temporal coarse-to-fine mechanism; the evidence only supports 'coarse-to-fine plus multi-scale' versus 'single-scale.' The distinction matters because a coarse-to-fine curriculum is a specific optimization claim, while a multi-scale objective is a different, weaker claim. The proposed test (multi-scale fixed) would settle this. However, the paper's results are not obviously wrong, and the concern is a missing ablation rather than a demonstrated flaw, so the conditional verdict stands.","tokens_in":9295,"tokens_out":5524,"duration_ms":55821,"concrete_test":"Retrain D with Eq. 6 applied to all granularities at every epoch (sum over K = 1, 2, 4, ..., 256), keeping the same architecture, epochs, optimizer, and loss budget. If this fixed multi-scale model reaches D+M's Table 1 scores within a standard error, the coarse-to-fine schedule is not necessary. A second arm in reverse order (fine-to-coarse) would further test whether ordering matters. For G+M+P, compare against static feature matching: compute L1 on layers 9, 7, 5, and 3 from epoch 1 with the same coefficient 100. If static matches dynamic, the 'dynamic' temporal schedule is not the source of improvement. Run at least three seeds to estimate variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 reports D+M beating D, and the paper attributes this gain to optimizing cosine similarity loss from coarse to fine granularity. However, the comparison only opposes a single-scale loss to a schedule that switches granularity every 20 epochs (2^14 down to 2^6). A model trained with the same multi-scale objective but with all granularities active at every epoch is absent. If that model performs as well as D+M, the temporal ordering is not the causal ingredient; the improvement is explained by the multi-scale objective alone. The same confound affects the generative results: G+M has no fixed multi-scale control, and G+M+P is introduced in Section 4.3 by adding L1 feature losses at layers 9, 7, 5, and 3 every 80 epochs, with no static all-layers feature-matching baseline. Additionally, many Table 1 differences are small (e.g., CBAK 3.35 vs 3.32, COVL 3.33 vs 3.20), and no error bars or repeated-seed statistics are provided, so the schedule-specific claim is not robustly supported. This is a missing-control issue, not an internal inconsistency; the reported numbers may be correct, but they do not demonstrate the claimed mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a coarse-to-fine optimization schedule for the cosine similarity loss in speech enhancement. The enhancement network is trained first with the loss computed on entire utterances and then on progressively shorter segments, down to a fine granularity; the same schedule is applied to the cosine regularization term of a conditional GAN, and a so-called dynamic perceptual loss is added by matching discriminator features at layers 9, 7, 5, and 3 at later stages of training. Experiments on the Valentini dataset compare a discriminative model D (the authors' reimplementation of cRM Rn), D with coarse-to-fine loss (D+M), a GAN baseline (G), G with coarse-to-fine regularization (G+M), and G+M with dynamic perceptual loss (G+M+P). The reported results show consistent improvements on CSIG, CBAK, COVL, and PESQ for D+M over D and for G+M+P over G+M, with the best generative result being CSIG 4.00.","tokens_in":9653,"tokens_out":3630,"duration_ms":38391,"significance":"If the causal claim is correct, the contribution is useful and general: a training-schedule modification that improves both discriminative and generative speech enhancement without changing the architecture or inference procedure. The evaluation is grounded in an external public benchmark with standard objective metrics, and the loss is not constructed from those metrics, so the gains are not circular. The authors also provide a reasonable set of ablations for the generative model (G, G+M, G+M+P) and disclose that the original cRM Rn numbers had a technical error, which is honest and source-credible. The significance is tempered by the fact that the core mechanism, temporal coarse-to-fine scheduling, is not isolated from a plain multi-scale objective, and by the absence of error bars around small metric differences.","major_comments":[{"comment":"The central claim that the improvement of D+M over D comes from the temporal coarse-to-fine schedule is not isolated. D+M uses a schedule that switches from granularity 2^14 to 2^6 by halving every 20 epochs, while the comparison D uses a single granularity. There is no control trained with all granularities active at every epoch, so the improvement could be due to the multi-scale objective alone rather than to the ordering of scales. A control that averages Eq. (6) over all Ki at each epoch is required to support the stated mechanism; the same control is needed for G+M.","section":"§4.2, Table 1"},{"comment":"The dynamic perceptual loss (DPL) is underspecified and the claimed benefit of making the feature matching dynamic is not tested. The text says that G+M+P 'start[s] with the original L2 loss ... and every 80 epochs' computes L1 loss on features at layers 9, 7, 5, 3, but it does not state whether layers are added progressively or all at once at the first switch, nor what the feature-loss schedule is after each 80-epoch milestone. Without a static baseline that applies the same multi-layer feature matching from the start of training, the results cannot distinguish the effect of the coarse-to-fine scheduling of the perceptual loss from the effect of adding feature matching at all.","section":"§4.3, Eq. (7)"},{"comment":"The experimental comparisons are single-run and no error bars or repeated-seed statistics are reported. Several of the decisive differences are small, for example CBAK 3.35 vs. 3.32 for D+M vs. D, COVL 3.33 vs. 3.31 for G+M+P vs. G+M, and SSNR 9.40 vs. 9.90 where D+M is worse than the single-scale baseline. Because the claims are quantitative and partly hinge on differences of 0.01-0.08, the paper should report means and variances over at least three seeds and, ideally, paired significance tests on the 824 test utterances.","section":"Table 1, Figures 3 and 4"},{"comment":"The schedule hyperparameters are introduced without sensitivity analysis: halving the granularity every 20 epochs, the coefficient 40 for the cosine regularization term, the coefficient 100 for the perceptual loss, the choice of layers 9, 7, 5, 3, and the 80-epoch interval for the perceptual loss are all presented as fixed choices. Since the paper proposes a general optimization strategy, it should at least show that the improvements are not critically dependent on these specific values, or provide a principled justification for them.","section":"§4.2, §4.3"}],"minor_comments":[{"comment":"The last paragraph of Section 4.3 refers to 'G+M+D' where the context and Table 1 indicate the intended model is 'G+M+P'; please correct this typo.","section":"§4.3"},{"comment":"The caption states that the red blocks are layers 9, 7, 5, 3 'from top to bottom,' but the figure shows an encoder-decoder without explicit layer indices; please clarify whether the numbering is from the encoder input or from the decoder output.","section":"Fig. 1"},{"comment":"The text says the loss is computed on the waveform following inverse STFT, but the network output is a complex-valued mask in the frequency domain; please state explicitly that the waveform is reconstructed via differentiable inverse STFT during training so that the gradient flows back to the mask.","section":"§2.2"},{"comment":"There are several reference formatting errors, including 'Nerual Information Processing' in [16] and the truncated URL-only entries for [5] and [17]; please use a consistent bibliography style and verify titles and venues.","section":"References"},{"comment":"The motivation that higher-dimensional vectors admit 'more feasible solutions' with the same cosine loss is only heuristic; a short formal argument or a reference to the geometry of the unit sphere would make the motivation more precise.","section":"§2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper reads like an extended abstract, and the missing multi-scale control is the main substantive gap. I would not reject it: the benchmark, baselines, and ablations are appropriate, and the reported improvements are consistent across most metrics. However, the load-bearing claim about temporal coarse-to-fine scheduling needs the fixed-multi-scale control, and the small metric differences need variance information, before the result can be accepted as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The paper proposes a coarse-to-fine training schedule for cosine similarity loss in speech enhancement, and a 'dynamic perceptual loss' for GANs. The core idea is borrowed from vision: start optimizing the loss on long segments and halve the segment length every 20 epochs. The claimed payoff is about 0.1–0.2 on CSIG/COVL/PESQ over a single-scale baseline on the Valentini benchmark.\n\nWhat is genuinely new is the application of this schedule to speech enhancement, plus the GAN extension where feature-matching layers are added over time. The experiments are run on a standard benchmark, include both discriminative and generative settings, and have clear ablations (D vs D+M, G vs G+M vs G+M+P). They also report SSNR where the method is worse and give a plausible explanation. That is honest.\n\nThe soft spots are real but not fatal. The main claim is that the temporal ordering—coarse to fine—is what helps. The paper never trains a model with all granularities active at every epoch. So the improvement could come from the multi-scale objective alone, not the schedule. The same confound applies to dynamic perceptual loss: there is no static all-layers feature-matching baseline. Second, all results are single-run, no error bars or seed variation, and some differences are small (CBAK 3.35 vs 3.32). Third, the schedule is underspecified: halving every 20 epochs and coefficients 40/100 are given without sensitivity analysis. These look hand-picked. None of this means the reported numbers are wrong; it means the mechanism claim is not established.\n\nWho is this for? People working on training tricks for audio enhancement or audio GANs. A reader wanting a quick empirical result on Valentini will get something. A reader wanting to understand why coarse-to-fine works will be left unsatisfied.\n\nMy recommendation: send it to review, not desk reject. The missing-control issue is fixable and should be the main request: add a static multi-scale control and multiple seeds. With those, it would be a solid incremental contribution.","headline":"Useful but under-controlled empirical claim: no static multi-scale baseline, so the coarse-to-fine mechanism is unproven.","tokens_in":10071,"tokens_out":2512,"would_cite":false,"duration_ms":22111,"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 that scheduling cosine similarity loss from coarse to fine granularities improves speech enhancement in discriminative and generative models, and that dynamic perceptual loss yields the best generative results.","keywords":["speech enhancement","coarse-to-fine optimization","cosine similarity loss","dynamic perceptual loss","generative adversarial network","discriminative model","generative model","time-frequency masking"],"falsifier":"Train the same network on the same dataset with a loss that at every epoch sums the cosine similarity loss over all granularities from whole clip to 64-sample chunks, or randomizes the order of granularities. If that model matches the coarse-to-fine model's CSIG, COVL, and PESQ scores, the specific temporal schedule is not the cause of the improvement.","tokens_in":9130,"feed_emoji":"🎙️","tokens_out":6824,"duration_ms":60237,"temperature":0.7,"pith_summary":"The paper argues that the way a speech-enhancement loss is sliced during training matters: instead of comparing the full enhanced waveform with the clean one as a single high-dimensional vector, the model should first be trained on the whole signal and then, every 20 epochs, on halves, quarters, and eventually 64-sample chunks of the cosine similarity loss. The reason is that cosine similarity computed on a long vector is too loose: many wildly different outputs can share the same cosine score, so optimizing on progressively finer segments adds constraints that make the predicted speech resemble the clean speech more closely. The paper shows this coarse-to-fine schedule improves a state-of-the-art discriminative network and a conditional GAN generator on perceived-quality metrics such as CSIG, COVL, and PESQ. It also proposes dynamic perceptual loss, which applies the same coarse-to-fine idea to the discriminator's adversarial loss by comparing intermediate deep features at coarser to finer layers over training, and reports the best generative-model score in its comparison table.","feed_headline":"Coarse-to-fine loss schedule lifts speech enhancement scores","feed_subtitle":"From whole clips down to 64-sample chunks, coarse-to-fine cosine loss lifts quality in both model types.","key_machinery":"The core mechanism is the slice-wise cosine similarity loss defined in Eq. 6: the signal is divided into $K_i$ vectors of dimension $g_i = \\text{dim}(x)/K_i$, and the loss averages cosine similarity over those slices. Training begins with $K_1 = 1$ (the whole $2^{14}$-sample clip), then halves the slice size every 20 epochs until $g_i = 2^6$, so each finer granularity places additional angle constraints on sub-segments of the prediction. For the GAN, dynamic perceptual loss extracts feature maps at layers 9, 7, 5, and 3 of the siamese discriminator and compares the fake and real utterances with L1 at progressively finer resolutions, adding a coarse-to-fine term to the adversarial loss.","core_discovery":"The paper's central claim is that scheduling cosine-similarity loss from coarse to fine granularity improves speech enhancement over a fixed single-granularity loss in both discriminative and generative settings. Concretely, the coarse-to-fine discriminative model reaches CSIG 3.94, CBAK 3.35, COVL 3.33, and PESQ 2.73, versus 3.79, 3.32, 3.20, and 2.62 for the single-scale baseline, while SSNR drops from 9.90 to 9.40. For the GAN, adding dynamic perceptual loss on top of coarse-to-fine training reaches CSIG 4.00, the best among the generative models compared, with COVL 3.34 and PESQ 2.69. The paper attributes these gains to the added constraints from optimizing at multiple granularities over the course of training, and argues the strategy generalizes to any objective whose granularity can be varied, including adversarial losses.","pith_inferences":["The paper never runs a control that uses all granularities at every epoch. If such a control matches or beats the schedule, the causal claim about temporal ordering would collapse to multi-scale supervision; the schedule's order would then be a convenience, not the mechanism.","The schedule parameters (halving every 20 epochs, coefficients 40 and 100) are chosen for the fixed 180-epoch protocol; for other datasets one could make the granularity switch adapt on loss stagnation or validation quality, which would be a direct test of whether the schedule, not just multi-scale loss, is what helps.","The same granularity-slicing logic applies to cosine losses used outside audio, such as image or embedding similarity; a testable extension is to replace a single global cosine loss with this schedule on a vision reconstruction task and look for the same quality gains.","The drop in SSNR under the discriminative coarse-to-fine model suggests the method trades exact waveform-level signal-to-noise for perceptually motivated scores; an ablation measuring listening tests or automatic speech recognition word error rate would say whether the trade is worthwhile."],"forward_implications":["Coarse-to-fine training transfers to any model whose loss is computed on the full sequence: the paper demonstrates it on a 20-layer discriminative network and on a conditional GAN generator.","Perceived-quality metrics that correlate with subjective listening improve under the schedule, while SSNR, which the paper notes correlates poorly with overall quality, gets worse.","Dynamic perceptual loss gives the best generative results in the comparison, suggesting that intermediate discriminator features at multiple resolutions are a usable adversarial training signal.","Because dynamic perceptual loss is computed from the siamese discriminator's internal features, it applies to any conditional GAN whose discriminator is built as a siamese network.","The improvement is not confined to a single final checkpoint: Figures 3 and 4 show coarse-to-fine and dynamic perceptual loss track above single-scale training throughout training."],"supporting_citations":[{"why":"Supplies the cosine similarity loss that the paper slices into different granularities.","marker":"[1]"},{"why":"The SEGAN generative baseline and source of the GAN training strategy and audio slicing convention.","marker":"[4]"},{"why":"State-of-the-art generative baseline with deep feature losses whose open-source numbers are re-run for comparison.","marker":"[5]"},{"why":"Provides the cRM Rn network reused as the discriminative model and the GAN generator; the paper notes a technical error in the original reported accuracy.","marker":"[8]"},{"why":"MMSE-GAN, a time-frequency masking generative baseline included in the comparison table.","marker":"[25]"},{"why":"WaveNet-based generative baseline included in the comparison table.","marker":"[30]"},{"why":"The Valentini noisy speech dataset used for all training and evaluation.","marker":"[31]"},{"why":"Defines the CSIG, CBAK, and COVL objective quality metrics used in all comparisons.","marker":"[34]"},{"why":"Defines the PESQ metric used for perceptually weighted speech quality evaluation.","marker":"[35]"}],"fun_headline_variants":["Coarse-to-fine cosine loss sharpens enhanced speech quality","Multiscale loss schedule boosts speech enhancement in both models","Dynamic perceptual loss refines GAN speech enhancement results","Training cosine loss from coarse to fine improves speech clarity","Coarse-to-fine optimization lifts speech enhancement metrics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the temporal coarse-to-fine ordering itself, rather than simply having multiple granularities at every epoch, explains the improved quality; the experiments do not include the control needed to separate the two.","fun_headline_variants_meta":{"raw":{"variants":["Coarse-to-fine cosine loss sharpens enhanced speech quality","Multiscale loss schedule boosts speech enhancement in both models","Dynamic perceptual loss refines GAN speech enhancement results","Training cosine loss from coarse to fine improves speech clarity","Coarse-to-fine optimization lifts speech enhancement metrics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000353,"raw_usage":{"total_tokens":1916,"prompt_tokens":934,"completion_tokens":982,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":903}},"tokens_in":550,"tokens_out":982,"duration_ms":9703,"temperature":1.0,"reasoning_tokens":903,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:50:53.414015+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same network on the same dataset with a loss that at every epoch sums the cosine similarity loss over all granularities from whole clip to 64-sample chunks, or randomizes the order of granularities. If that model matches the coarse-to-fine model's CSIG, COVL, and PESQ scores, the specific temporal schedule is not the cause of the improvement.","supporting_citations":[{"cited_title":"It has quite a few applications including noise cancelling, audio editing, prepro- cessing for speech recognition, just to name a few","cited_arxiv_id":null,"evidence_quote":"Supplies the cosine similarity loss that the paper slices into different granularities."},{"cited_title":"Particularly in generative models, the proposed dynamic per- ceptual loss could further improve the accuracy obtained from optimizing the cosine similarity loss from coarse to fine","cited_arxiv_id":null,"evidence_quote":"The SEGAN generative baseline and source of the GAN training strategy and audio slicing convention."},{"cited_title":"Inspired by the coarse-to-fine idea, we further pro- posed the dynamic perceptual loss as the adversarial loss term in the generator training of GAN","cited_arxiv_id":null,"evidence_quote":"State-of-the-art generative baseline with deep feature losses whose open-source numbers are re-run for comparison."},{"cited_title":"Enhancement of speech corrupted by acoustic noise,","cited_arxiv_id":null,"evidence_quote":"Provides the cRM Rn network reused as the discriminative model and the GAN generator; the paper notes a technical error in the original reported accuracy."},{"cited_title":"Deep multi-scale video prediction beyond mean square error,","cited_arxiv_id":null,"evidence_quote":"MMSE-GAN, a time-frequency masking generative baseline included in the comparison table."},{"cited_title":"Short term spectral analysis, synthesis, and mod- ification by discrete fourier transform,","cited_arxiv_id":null,"evidence_quote":"WaveNet-based generative baseline included in the comparison table."},{"cited_title":"Time-frequency masking in the complex domain for speech dereverberation and denoising,","cited_arxiv_id":null,"evidence_quote":"The Valentini noisy speech dataset used for all training and evaluation."},{"cited_title":"Rectifier nonlinear- ities improve neural network acoustic models,","cited_arxiv_id":null,"evidence_quote":"Defines the CSIG, CBAK, and COVL objective quality metrics used in all comparisons."},{"cited_title":"Deepface: Closing the gap to human-level performance in face verifica- tion,","cited_arxiv_id":null,"evidence_quote":"Defines the PESQ metric used for perceptually weighted speech quality evaluation."}],"review_version":1}