{"id":"c6f3c6e8-2de7-42bb-8905-76e191401f83","arxiv_id":"2411.12268","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-stage neural vocoder predicts noisy amplitude and phase spectra from a noisy mel-spectrogram, denoises them with MP-SENet, and reconstructs clean speech, outperforming HiFi-GAN and Vocos on VoiceBank+DEMAND.","lead":"This paper builds a denoising vocoder that turns a noisy mel-spectrogram directly into clean audio. It combines a spectrum predictor with an enhancement model and reports strong benchmark scores.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"STFT mismatch is argued away by the paper's own ablation; the load-bearing gap is that 'state-of-the-art' rests on only two vocoder baselines, no significance testing, and an undisclosed frozen vs. fine-tuned enhancement module.","rationale":"The reader's weakest_assumption (MP-SENet checkpoint generalization under STFT mismatch) is real but the paper's own analysis in Section 3.4 substantially defuses it: a deliberately matching STFT configuration yields much worse numbers, so the mismatched cascade is not an accidental failure mode but a defended design choice. My concern shifts to the empirical evidence for state-of-the-art: too few baselines (especially no DNR-HiNet), no statistical testing, and an ambiguity about whether the enhancement module is frozen or fine-tuned. None of this undermines the architecture's plausibility or the novelty of using a noisy spectral bridge, so the correct verdict remains CONDITIONAL rather than ACCEPT or REJECT. The concrete test I propose (add DNR-HiNet and CIs; disclose/freeze-or-finetune) directly targets the remaining load-bearing uncertainty.","tokens_in":6303,"tokens_out":3193,"duration_ms":28725,"concrete_test":"Benchmark the proposed system against DNR-HiNet (with and without its F0/mel input variant adapted to mel-only) on the VoiceBank+DEMAND test set, and report 95% confidence intervals or paired significance tests (e.g., Wilcoxon) over the 872 clips for PESQ/STOI/COVL. In the same experiment, disclose whether the MP-SENet module is frozen or fine-tuned and rerun the best configuration both ways. If DNR-HiNet is not available, at minimum re-run Table 1 with bootstrap CIs: if overlaps occur with HiFi-GAN on STOI (0.92 vs. 0.94), the 'significantly outperformed' claim needs qualification.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the proposed two-stage denoising vocoder is state-of-the-art among neural vocoders. The paper's Table 3 directly addresses the STFT mismatch concern: matching MP-SENet's STFT configuration in the spectrum predictor drops PESQ from 2.88 to 2.26 and SSNR to -0.31, so the mismatched cascade is an empirically defended design choice. The real load-bearing gap is the comparison set and protocol. (1) The only vocoder baselines are HiFi-GAN and Vocos modified into denoisers; DNR-HiNet, the closest prior denoising vocoder (F0+mel inputs), is cited but never evaluated, so the SOTA claim lacks its most direct competitor. (2) The SE comparisons are MetricGAN and MetricGAN+, both amplitude-only and from 2019/2021, not current SOTA. (3) No error bars, no significance tests, and a single test set support the word 'significantly.' (4) The training recipe is under-specified: the paper says MP-SENet training follows [9], but does not state whether the open-source checkpoint is frozen, fine-tuned, or co-trained with the spectrum predictor during the 700k steps. If MP-SENet is frozen, the numbers depend on how well that checkpoint's internal feature statistics tolerate the 1024-point-FFT spectra; if fine-tuned, the reported recipe omits a potentially critical component. These gaps make the SOTA claim plausible but not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a neural denoising vocoder that maps a noisy mel-spectrogram to a clean waveform in two stages. A spectrum predictor (an APNet-derived module with separate amplitude and phase predictors) first estimates the noisy amplitude and phase spectra from the noisy mel-spectrogram; an enhancement module (MP-SENet) then denoises these spectra; the clean waveform is reconstructed by iSTFT. Experiments on VoiceBank+DEMAND compare against HiFi-GAN and Vocos as vocoder baselines and against MetricGAN/MetricGAN+ as speech-enhancement baselines, with additional ablations that replace the spectrum predictor with the original APNet and that match the STFT configuration of MP-SENet. The paper claims state-of-the-art performance among neural vocoders and performance comparable to several speech-enhancement methods.","tokens_in":6596,"tokens_out":4806,"duration_ms":49192,"significance":"If the results hold, the paper demonstrates that a mel-only denoising vocoder is practical and that a two-stage spectral bridge (noisy amplitude/phase prediction followed by spectral enhancement) is an effective design. The work builds on existing components (APNet and MP-SENet) and is clearly motivated by real noisy TTS scenarios. The paper includes a useful ablation: Table 3 shows that deliberately matching the STFT configuration of MP-SENet degrades PESQ from 2.88 to 2.26, which is an empirical defense of the mismatched-configuration design. Equation (1) is a true identity and is used only as motivation, so there is no circularity in the evaluation. The main weaknesses are in the evaluation protocol: the SOTA claim rests on only two vocoder baselines, no significance testing accompanies the metric comparisons, and the training/fine-tuning status of MP-SENet is ambiguous.","major_comments":[{"comment":"The central claim of state-of-the-art performance among neural vocoders is supported only by comparison with HiFi-GAN and Vocos, both of which are standard clean-speech vocoders that the paper modifies by training them on noisy mel-spectrograms. The manuscript itself identifies DNR-HiNet [7] as the closest prior denoising vocoder, but DNR-HiNet is never evaluated. A SOTA claim requires either an evaluation against the most relevant existing denoising vocoder or an explicit and justified statement of why that comparison is excluded. As written, the claim overreaches the evidence.","section":"Section 3.2, Table 1"},{"comment":"The manuscript does not specify whether the open-source MP-SENet checkpoint is kept frozen, fine-tuned, or jointly trained with the spectrum predictor during the 700k training steps. The text says both 'we use the open source checkpoint of MP-SENet' and 'Training of all models employed AdamW ... for 700k steps,' which is ambiguous. This is load-bearing because the enhancement module receives spectra computed with a 1024-point FFT that differ from the 400-point FFT configuration used when the MP-SENet checkpoint was trained, and the behavior of the cascade depends critically on whether the enhancement module is adapted to this input distribution. Please state explicitly which parameters are updated in each training stage.","section":"Section 3.1, Implementation"},{"comment":"All objective metrics are reported as point estimates on a single test set, without confidence intervals, significance tests, or multiple training runs. The word 'significantly' in Section 3.2 is therefore not statistically supported. While the gaps over HiFi-GAN and Vocos appear large, the comparison with MetricGAN+ in Table 2 is close (e.g., PESQ 2.88 vs 3.15, CBAK 3.06 vs 3.16), and the claim of 'comparable' performance cannot be assessed without error bars. The authors should report variance or significance, or at minimum temper the language of 'significantly.'","section":"Section 3.2, Tables 1-3"},{"comment":"The speech-enhancement comparisons are limited to MetricGAN (2019) and MetricGAN+ (2021), which are not representative of current state-of-the-art SE systems. The abstract's claim of 'comparable performance with the serveral advanced speech enhancement methods' is therefore weakened, since the selected baselines are more than three years old and are amplitude-only methods. Either update the SE baseline set to more recent methods or narrow the claim to 'comparable to the evaluated SE methods.'","section":"Section 3.3, Table 2"}],"minor_comments":[{"comment":"There are several typos: 'serveral' in the abstract, 'spectrum predictior' in Section 2.1, 'metircs' in Section 3.1, 'T able' in table captions, 'SE filed' in Section 2.3, and 'MeticGAN+' in Section 3.3.","section":"Abstract, Section 2.1, Section 3.1"},{"comment":"Training details are incomplete: the manuscript gives the learning rate and number of steps but not the learning-rate schedule, batch size, number of GPUs, or whether the spectrum predictor and enhancement module are trained separately or jointly. These details are needed for reproducibility.","section":"Section 3.1, Implementation"},{"comment":"The STFT-mismatch ablation is commendable and directly addresses a likely reviewer concern, but it is also a single-run point estimate. Reporting variance or a small number of runs would make the conclusion that the mismatched configuration is preferable more robust.","section":"Section 3.4, Table 3"},{"comment":"Table 2 omits SSNR and STOI because the original MetricGAN papers did not report them, but these metrics are standard on VoiceBank+DEMAND. If released checkpoints are available, computing them for MetricGAN/MetricGAN+ would provide a more complete comparison; otherwise the omission should be noted in the table caption.","section":"Section 3.3, Table 2"},{"comment":"The enhancement module is described as using MP-SENet with Conformer blocks, amplitude mask decoder, and phase decoder. A few sentences on how the phase decoder is trained (e.g., the anti-wrapping loss) would help readers who do not know MP-SENet.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own prior systems (APNet, APNet2, MP-SENet, DNR-HiNet), which is not inherently a problem, but it makes the missing DNR-HiNet comparison and the ambiguous MP-SENet training status more consequential. I would also note that the paper's length and evaluation depth are more typical of a workshop or conference paper; for a journal, the requested strengthening of the comparison set and statistical protocol is necessary. The audio demo link is useful, but releasing the code or a recipe would considerably increase the value of the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is a mel-only denoising vocoder: the spectrum predictor converts the noisy mel into noisy amplitude/phase, and MP-SENet cleans those spectra before iSTFT. That removes the F0 requirement of the earlier DNR-HiNet, which is a real step forward for noisy TTS and VC pipelines. The paper is honest about what it does and does not achieve — it positions the method as comparable to, not better than, the SE baselines on several metrics, and the ablation in Table 3 actually defends the STFT mismatch, since forcing the spectrum predictor to match MP-SENet's STFT parameters drops PESQ from 2.88 to 2.26. I think the stress-test note has the right emphasis: the STFT concern is largely answered by the paper itself, and the load-bearing gap is the comparison set and the training protocol.\n\nWhat the paper does well: the architecture is clean, the equation motivating the phase bridge is a true identity, and the ablations are informative. The reported gains over HiFi-GAN and Vocos, modified to take noisy mel input, are consistent across all six metrics. The audio demo page is a good touch, and the paper does not oversell its SE comparison.\n\nThe soft spots, in proportion: first, the \"state-of-the-art performance compared to existing neural vocoders\" claim rests on exactly two baselines, and DNR-HiNet — the closest prior denoising vocoder — is cited but never evaluated. That is the main hole. Second, no error bars or significance tests, so \"significantly outperformed\" is just a word. Third, the MP-SENet training recipe is genuinely ambiguous: the paper says it uses the open-source checkpoint, but also says training all models ran for 700k steps with a 2e-4 learning rate. Readers cannot tell whether MP-SENet is frozen, fine-tuned, or co-trained. That matters because the cascade's performance could depend on how much adaptation happens. None of these flaws are fatal — the contribution is a plausible engineering result — but they keep the strongest claims from being fully established.\n\nWho this is for: people working on noise-robust TTS or voice conversion who want a vocoder that runs from mel alone, and anyone building on APNet/MP-SENet-style spectral pipelines. The paper deserves a serious referee: the idea is novel enough and the evidence is suggestive enough to warrant a careful look, with the expectation that the authors tighten the comparison and clarify the training protocol before acceptance.","headline":"A useful mel-only denoising vocoder with a defensible two-stage design; the SOTA claim needs a bigger comparison set and a stated training recipe.","tokens_in":7158,"tokens_out":1683,"would_cite":true,"duration_ms":20176,"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 two-stage denoising vocoder generates clean speech directly from noisy mel-spectrograms.","keywords":["neural vocoder","denoising vocoder","speech enhancement","amplitude spectrum prediction","phase spectrum prediction","mel-spectrogram","inverse short-time Fourier transform","VoiceBank+DEMAND"],"falsifier":"Ablate the spectrum predictor by feeding the noisy mel-spectrogram straight into the enhancement module and measuring PESQ and STOI on VoiceBank+DEMAND. If the simplified system matches the full pipeline, the claim that noisy spectra must be predicted before enhancement is not supported.","tokens_in":6077,"feed_emoji":"🎙️","tokens_out":6675,"duration_ms":63296,"temperature":0.7,"pith_summary":"This paper claims that a neural vocoder can be turned into a denoising vocoder by splitting the job into two frame-level spectral steps. First, a spectrum predictor reconstructs the noisy amplitude and phase spectra that the input noisy mel-spectrogram has lost. Then an enhancement module cleans those spectra before an inverse short-time Fourier transform rebuilds the waveform. On the VoiceBank+DEMAND test set, the system outperforms two standard neural vocoders, HiFi-GAN and Vocos, and is comparable to two established speech-enhancement models that receive far richer amplitude-spectrum input.","feed_headline":"Two-stage vocoder turns noisy mel-spectrograms into clean speech","feed_subtitle":"Two-stage spectral prediction does it without phase information, beating standard vocoders.","key_machinery":"The load-bearing mechanism is the two-stage cascade in which noisy amplitude and phase spectra serve as an explicit intermediate representation. The first stage, a spectrum predictor adapted from the APNet vocoder, uses separate ResNet-based amplitude and phase predictors to convert the noisy mel-spectrogram into full noisy spectra; the phase predictor matters most because phase information is entirely absent from the mel input. The second stage is the MP-SENet enhancement module, a conformer-based model with separate amplitude-mask and phase decoders that turns the noisy spectra into clean ones, followed by iSTFT to synthesize the waveform. Training the spectrum predictor uses a multi-resolution discriminator and linear anti-wrapping phase losses, while the enhancement module follows its original time-domain, amplitude, complex-spectrum, and anti-wrapping losses with adversarial training.","core_discovery":"The central claim is that clean waveforms can be recovered from a noisy mel-spectrogram by explicitly predicting the noisy amplitude and phase spectra first and then denoising them, rather than asking a single network to jump directly from mel input to clean speech. The authors argue that direct clean-phase recovery from the mel-spectrogram is especially fragile at low SNR because the noisy phase is the clean phase plus an angle term that grows as the noise-to-signal ratio rises. Their two-stage design uses the predicted noisy spectra as a bridge: the spectrum predictor supplies what the mel-spectrogram is missing, and the enhancement module, built on a parallel magnitude-and-phase denoising model, removes the residual noise. Experiments on VoiceBank+DEMAND show the full system reaching a PESQ of 2.88 and STOI of 0.94, surpassing the two vocoder baselines and landing between MetricGAN and MetricGAN+ on the composite quality scores.","pith_inferences":["Because the enhancement module is used as a frozen checkpoint with different STFT settings from the spectrum predictor, the reported gains may partly reflect how well that checkpoint generalizes across feature configurations; a version fine-tuned end-to-end on the bridge features would separate this transfer effect from the architecture's benefit.","The same two-stage bridge could be tested with other lossy acoustic inputs, such as speaker embeddings or noise-robust features, or with other enhancement backbones, to see whether the noisy-spectrum bridge is the general principle.","The phase-only angle equation suggests a direct extension: a predictor that estimates the noise-to-signal ratio per time-frequency bin could let the system condition its denoising on expected phase deviation, potentially improving very-low-SNR cases."],"forward_implications":["A vocoder that accepts only mel-spectrograms can perform denoising, so noisy speech from a target speaker can be converted to clean synthesis without requiring a separate waveform-enhancement front end.","Predicting noisy spectra as an intermediate beats trying to predict clean spectra directly, especially for phase recovery in low-SNR conditions.","Off-the-shelf speech-enhancement models can be recycled as the second stage of a denoising vocoder, suggesting a modular route to other vocoder–enhancer combinations.","The system reaches the quality level of amplitude-input speech-enhancement models despite operating from a lossy mel-spectrogram, setting a practical baseline for mel-only denoising."],"supporting_citations":[{"why":"Supplies the amplitude/phase spectrum predictor design the first stage adapts.","marker":"[8]"},{"why":"Supplies the enhancement module used as the second stage.","marker":"[9]"},{"why":"Prior denoising vocoder that needed F0 input; the proposal removes that requirement.","marker":"[7]"},{"why":"One of the two vocoder baselines that fail on noisy mel input.","marker":"[1]"},{"why":"The other vocoder baseline, a Fourier-based vocoder, also outperformed by the proposal.","marker":"[2]"},{"why":"Provides the anti-wrapping phase losses used in both stages.","marker":"[11]"},{"why":"Supplies the paired clean/noisy dataset and train/test splits for all comparisons.","marker":"[14]"},{"why":"MetricGAN baseline whose amplitude-spectrum input highlights the harder mel-only setting.","marker":"[13]"}],"fun_headline_variants":["Two-stage vocoder denoises mel-spectrograms to clean speech","From noisy mel-spectrogram to clean waveform via phase prediction","New vocoder predicts amplitude and phase to clean speech","Clean speech from noisy mel-spectrogram without phase info","Two-stage network recovers speech from noisy spectrograms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline leans on a pre-trained enhancement checkpoint, built for one audio feature configuration, working correctly on the different configuration produced by the spectrum predictor, even though the two stages were never jointly trained or explicitly aligned.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage vocoder denoises mel-spectrograms to clean speech","From noisy mel-spectrogram to clean waveform via phase prediction","New vocoder predicts amplitude and phase to clean speech","Clean speech from noisy mel-spectrogram without phase info","Two-stage network recovers speech from noisy spectrograms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000665,"raw_usage":{"total_tokens":3024,"prompt_tokens":921,"completion_tokens":2103,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":2022}},"tokens_in":537,"tokens_out":2103,"duration_ms":15317,"temperature":1.0,"reasoning_tokens":2022,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:43:50.780570+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Ablate the spectrum predictor by feeding the noisy mel-spectrogram straight into the enhancement module and measuring PESQ and STOI on VoiceBank+DEMAND. If the simplified system matches the full pipeline, the claim that noisy spectra must be predicted before enhancement is not supported.","supporting_citations":[{"cited_title":"IEEE/ACM Transactions on Audio, Speech, and Language Processing (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the amplitude/phase spectrum predictor design the first stage adapts."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Supplies the enhancement module used as the second stage."},{"cited_title":"IEEE/ACM Transac- tions on Audio, Speech, and Language Processing30, 2036–2048 (2022)","cited_arxiv_id":null,"evidence_quote":"Prior denoising vocoder that needed F0 input; the proposal removes that requirement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"One of the two vocoder baselines that fail on noisy mel input."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Provides the anti-wrapping phase losses used in both stages."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Supplies the paired clean/noisy dataset and train/test splits for all comparisons."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"MetricGAN baseline whose amplitude-spectrum input highlights the harder mel-only setting."}],"review_version":1}