{"id":"6e27ae33-dfc1-445e-b2d5-6a31281b119a","arxiv_id":"2505.19576","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Mel-McNet performs online multichannel speech enhancement in the Mel domain, reducing FLOPs by roughly 60% versus McNet while keeping speech quality and ASR accuracy comparable.","lead":"This paper proposes Mel-McNet, a multichannel speech enhancement model that processes audio in Mel frequency instead of the usual linear scale, cutting computation by about 60% with similar quality. It suggests that Mel-scale processing can make real-time speech enhancement more efficient on edge devices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Vocoder asymmetry confounds the SOTA claim: Mel-McNet is compared against waveform-output baselines without a McNet+Vocos control, so the DNSMOS/WER gains cannot be cleanly attributed to Mel-domain processing.","rationale":"The reader's weakest assumption—the missing McNet+Vocos baseline—is the same load-bearing concern I identify. It directly controls the paper's central empirical claim of 'outperforming other SOTA online methods' while leaving the efficiency claim intact. The concern is not about internal inconsistency: the FLOPs arithmetic is credible, and the STFT-to-Mel compression provides a structural reason for the 60% FLOPs reduction. However, the published Table 1 cannot distinguish Mel-scale processing gains from vocoder-induced perceptual scoring gains, and the paper's own 'Clean Mel+Vocos' upper-bound row demonstrates that the vocoder is not neutral across metrics. I also note the second-order asymmetry that Mel-McNet's LogMel can bypass vocoder reconstruction for ASR, while waveform baselines must re-extract features; this further weakens the WER comparison unless controlled. The proposed concrete test is the minimal experiment that settles the attribution question. Because this is precisely a conditional-acceptance concern rather than a fatal flaw, and because the central efficiency contribution is likely sound, I keep the reader's CONDITIONAL verdict unchanged rather than escalating or downgrading it.","tokens_in":8769,"tokens_out":5577,"duration_ms":59109,"concrete_test":"Run the missing McNet+Vocos control on the same CHiME-3 eval/test sets: (1) take McNet's enhanced time-domain outputs; (2) compute 80-band LogMel power spectrograms using the identical 512-sample/128-hop STFT and the same triangular mel filterbank as Mel-McNet; (3) pass these features through the exact same causal Vocos and evaluate DNSMOS/P.835, WB-PESQ, STOI, and WER with the same ASR; (4) for WER, also feed Mel-McNet's direct LogMel output and McNet's waveform-derived LogMel through the identical feature-extraction path. If McNet+Vocos matches or exceeds Mel-McNet's DNSMOS and WER numbers, the 'outperforms SOTA' claim is a vocoder artifact and only the FLOPs/efficiency claim survives. If Mel-McNet still leads, the Mel-domain processing contribution is validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 reports Mel-McNet DNSMOS 3.43 vs McNet 3.40, with WER tied at 4.4, and concludes Mel-McNet outperforms SOTA online methods. The comparison is asymmetric: McNet and the other baselines output time-domain waveforms, while Mel-McNet outputs an enhanced 80-band LogMel spectrogram that is reconstructed by the Vocos neural vocoder. The paper's own upper-bound row 'Clean Mel+Vocos' shows the vocoder materially changes evaluation outcomes—DNSMOS is preserved at 3.55 while WB-PESQ and STOI drop from 4.06/100.0 to 4.05/97.7—so the vocoder is not a neutral transform. Consequently, the 0.03 DNSMOS advantage over McNet (and the WER advantage over EaBNet and oSpatialnet-mamba) could be produced by the vocoder rather than by the Mel-scale STFT-to-Mel compression or backbone modification. The ablation in Table 2 only varies the compression module inside Mel-McNet, keeping the same Vocos backend for all rows; it cannot separate the Mel-domain contribution from the vocoder's effect. The 60% FLOPs reduction claim is structurally supported and not threatened by this concern, because that comparison is between enhancement networks (43.75 vs 115.13 G/s), but the paper's additional claim to 'outperform other SOTA' on CHiME-3 depends on an unfair or at least uncontrolled backend comparison. A second confound sits inside WER: Mel-McNet's LogMel features can be fed directly to the ASR model, whereas baselines must first synthesize a waveform and then re-extract the same features, a path known from the upper-bound row to be lossy. Both confounds are fixable with one control experiment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Mel-McNet, an online multichannel speech enhancement framework that compresses multichannel STFT features into an 80-band Mel-scale representation via a new STFT-to-Mel module and then processes spectral and spatial information with a modified McNet backbone operating directly in the Mel domain. The enhanced LogMel spectrogram can be fed either to the Vocos neural vocoder for waveform reconstruction or to a pretrained ASR system for transcription. Experiments on the CHiME-3 dataset report a roughly 60% reduction in FLOPs relative to the McNet backbone while maintaining comparable enhancement and ASR performance, and the paper claims to outperform other SOTA online methods. An ablation study compares alternative compression schemes for the STFT-to-Mel module.","tokens_in":9187,"tokens_out":5351,"duration_ms":44048,"significance":"If the efficiency result holds, the paper makes a useful contribution by demonstrating that Mel-scale processing can be applied to multichannel spatial information without a full-band recovery step, and the released code plus the use of an external FLOPs counter are strengths. The 60% FLOPs reduction is structurally supported by the compression from F=257 to F'=80 frequency bins and is robust even when the vocoder's FLOPs are included. However, the secondary claim of outperforming SOTA methods is weakened by a confounded comparison: Mel-McNet uses a neural vocoder while all baselines output waveforms directly, and the ASR evaluation is asymmetric. The paper's value would be significantly strengthened by adding a McNet+Vocos control and by reporting ASR results on a common feature-extraction path.","major_comments":[{"comment":"The claim that Mel-McNet 'outperforms other SOTA online methods' is not cleanly supported because the comparison changes two variables at once: the enhancement backbone and the output backend. Mel-McNet outputs an 80-band LogMel spectrogram that is converted to a waveform by the pretrained Vocos, whereas EaBNet, oSpatialnet-mamba, and McNet output time-domain waveforms directly. The 'Clean Mel+Vocos' row in Table 1 shows that Vocos is not a neutral transform for STOI (100.0 to 97.7) and WB-PESQ (4.06 to 4.05), even though DNSMOS is preserved. To attribute the DNSMOS improvement (3.43 vs 3.40) and the WER result (4.4) to Mel-scale processing, please add a 'McNet+Vocos' control: take McNet's enhanced waveform, extract its LogMel spectrogram, reconstruct with Vocos, and evaluate all metrics. Without this control, the observed advantages could be produced by the vocoder rather than by the STFT-to-Mel compression or the modified backbone.","section":"§3.2, Table 1"},{"comment":"The ASR evaluation is asymmetric. Mel-McNet's LogMel output is fed directly to the pretrained E-branchformer ASR, while all baselines must first synthesize a time-domain waveform and then extract LogMel features from that waveform. This direct path avoids possible mismatches in feature extraction and may favor Mel-McNet independent of enhancement quality. Please report a control in which all baselines' enhanced waveforms are converted to LogMel using the same feature pipeline and fed to the same ASR, or explicitly discuss this confound when interpreting the WER comparison.","section":"§2.2 and §3.2, Table 1"},{"comment":"The 'comparable performance' claim is not fully supported by STOI. Mel-McNet's STOI is 95.4% versus McNet's 97.7%, a 2.3-point drop, while WB-PESQ is tied at 3.99. The authors attribute the lower STOI to the vocoder, but the 'Clean Mel+Vocos' row shows that Vocos on clean LogMel yields 97.7% STOI, only 2.3 points below clean speech; it does not explain the additional 2.3-point drop to 95.4% in Mel-McNet. Please provide a decomposition (e.g., evaluate the enhanced LogMel before vocoding with a Mel-domain reference, or compare McNet+Vocos and Mel-McNet+Vocos on identical pipelines), or restrict the 'comparable' claim to the non-intrusive metrics and WER.","section":"§3.2, Table 1"}],"minor_comments":[{"comment":"The phrasing 'few works are implemented in a Mel-frequency domain' is awkward; consider 'few works operate in the Mel-frequency domain' or a similar wording.","section":"§1"},{"comment":"The notation in Eq. (1) uses the index i both for the block index in 'EMag_i' and 'EPha_i' and for the element-wise multiplication symbol; clarify the range of i (i = 1, ..., Q) and define the linear(·) layer's input/output dimensions.","section":"§2.1.1, Eq. (1)"},{"comment":"Please justify the choice of the fifth microphone as the reference channel, since the choice of reference can affect the reported enhancement and ASR metrics.","section":"§3.1"},{"comment":"The caption states that 'Param. and FLOPs of speech enhancement network/neural vocoder are given in addition' for Mel-McNet, but the table does not explicitly separate the two in the caption text; adding a note such as '(enhancement network + vocoder)' would improve readability.","section":"Table 1 caption"},{"comment":"The statement that EaBNet 'has lower computational complexity' while the ASR performance is 'heavily impacted' is presented without noting that EaBNet also has a different parameter count and architecture; the sentence would benefit from a brief context about the comparison being across different designs.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The core efficiency result (FLOPs reduction) appears sound and is worth publishing, but the paper overreaches in its SOTA claim. The required control experiments (McNet+Vocos and a common ASR feature path) are straightforward and should be requested before acceptance. I also note that the paper extends the authors' own McNet, which is appropriate but should be kept in mind in the editorial assessment; the manuscript does cite the previous work adequately."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline is: the 60% FLOPs reduction is real and structurally supported, but the 'outperforms SOTA' claim is not clean. The paper compares Mel-McNet, which outputs an 80-band LogMel spectrogram that a neural vocoder turns into a waveform, against McNet, EaBNet, and oSpatialnet-mamba, which output waveforms directly. Their own upper bound row, Clean Mel+Vocos, shows the vocoder is not neutral: WB-PESQ and STOI drop from 4.06/100.0 to 4.05/97.7 and MOS-BAK from 4.50 to 3.63. So the 0.03 DNSMOS advantage over McNet could just be the vocoder, not the Mel-scale compression. The WER comparison has a second asymmetry: Mel-McNet gives the ASR LogMel features directly, while the baselines must synthesize a waveform and re-extract features. One control row, McNet+Vocos, would fix both.\n\nWhat is actually new: the STFT-to-Mel module is a decent design. Separate branches for magnitude and phase, handcrafted filterbanks for magnitude, learnable ones for phase, attention-like interaction, and temporal smoothing. The Table 2 ablation shows each component pulls its weight, and the separate FCB+TSB variant wins. The FLOPs measurement is clean: 43.75 vs 115.13 G/s for the enhancement network, consistent with 257 to 80 bins. They retrain all baselines on the same data, use an external FLOPs counter, and released code. That is solid, reproducible work.\n\nThe soft spot is the evaluation asymmetry, and they half-acknowledge it in the upper-bound discussion but then claim SOTA anyway. That is a selective reading of their own table. The efficiency claim is not threatened, but the performance and SOTA claims need the control experiment.\n\nWho is this for: people working on real-time multichannel enhancement and Mel-domain representations. It is an incremental but useful data point that Mel compression cuts FLOPs substantially without wrecking quality. It does not change fundamentals.\n\nRecommendation: send it to peer review. The efficiency result deserves referee time. Ask for the McNet+Vocos baseline before accepting any performance claim.","headline":"Efficiency claim is solid and structurally supported, but the SOTA performance claim is confounded by the vocoder and ASR feature asymmetry; worth engaging for the FLOPs result, needs a control experiment.","tokens_in":9713,"tokens_out":2988,"would_cite":false,"duration_ms":26205,"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":"Mel-McNet claims that moving multichannel enhancement into the Mel-frequency domain cuts compute by nearly 60% while matching the original McNet's quality and ASR accuracy.","keywords":["multichannel speech enhancement","Mel-scale processing","online processing","McNet backbone","STFT-to-Mel compression","Mel power ratio mask","neural vocoder","CHiME-3"],"falsifier":"Run the McNet backbone through the same pipeline as Mel-McNet: convert McNet's enhanced STFT output to LogMel, reconstruct with the same Vocos vocoder, and evaluate DNSMOS and WER on CHiME-3. If this McNet plus Vocos baseline matches Mel-McNet's numbers, the 60% complexity reduction would be confirmed but the quality and ASR claim would not be due to Mel-scale processing.","tokens_in":8593,"feed_emoji":"🎙️","tokens_out":5939,"duration_ms":49625,"temperature":0.7,"pith_summary":"Mel-McNet is a framework for online multichannel speech enhancement that does almost all of its work on a Mel-frequency scale instead of the usual linear-frequency STFT scale. The paper's claim is that a compact STFT-to-Mel module plus a modified McNet backbone can generate enhanced LogMel spectrograms directly, and that these can be handed to a vocoder for waveform reconstruction or to an ASR system for transcription. On the simulated CHiME-3 benchmark the framework is reported to cut computational complexity by nearly 60% relative to the McNet backbone while keeping enhancement quality and word error rate comparable, and to beat other online multichannel methods on DNSMOS and WER. If true, this makes perceptually motivated frequency compression a viable design choice for real-time, low-complexity multichannel enhancement on edge devices.","feed_headline":"Mel-scale processing cuts speech enhancement compute by 60%","feed_subtitle":"A Mel-domain version of McNet keeps speech quality and ASR accuracy while running in real time.","key_machinery":"The load-bearing object is the STFT-to-Mel module: two parallel branches, one for multichannel magnitude and one for phase, each built from a linear projection, frequency compression blocks made of along-frequency 1-D convolutions and handcrafted or learnable Mel filterbanks, and a causal time-smoothing convolution. An attention-like information-communication operation lets the phase branch modulate the magnitude branch, and the compressed features are stacked and smoothed to produce a T frames by 80 Mel bands by D embedding for the McNet backbone. The training target is the rectified Mel power ratio mask, Mel-PRM, which clips the ratio of clean to noisy Mel power at 1. The module's role is to compress the frequency axis before the expensive along-frequency and along-time LSTM processing in McNet, which is where the nearly 60% complexity reduction comes from.","core_discovery":"The central claim is that the nonlinear Mel scale can carry the full multichannel enhancement computation without a performance penalty. The paper constructs Mel-McNet by inserting an STFT-to-Mel module before a modified McNet backbone, so that both spectral cues (magnitude) and spatial cues (phase) are compressed from F linear frequencies to 80 Mel bands before the four McNet modules process them. The network is trained to predict a rectified Mel power ratio mask (Mel-PRM), and at inference the enhanced LogMel power spectrogram is obtained by masking the noisy power and taking the logarithm. On the CHiME-3 evaluation the paper reports a 60% FLOPs reduction, a real-time factor of 0.799 on CPU, and DNSMOS/WER results that match or exceed the linear-frequency baselines; it attributes the small drops in WB-PESQ and STOI to the neural vocoder, not to the Mel-scale processing.","pith_inferences":["The reported gains are not yet isolated from the vocoder: since Mel-McNet outputs LogMel and reconstructs waveforms with Vocos while the linear-frequency baselines output waveforms directly, a McNet plus the same Vocos baseline would be needed to attribute the DNSMOS and WER results specifically to Mel-scale processing.","Because the CHiME-3 simulated training data is reverberation-free (time delay only), the framework's behavior under reverberation or with moving speakers is an open extension, not yet demonstrated.","The STFT-to-Mel front-end is modular and could be attached to other sub-band/full-band enhancement backbones; the paper names this as future work, and a direct test would be to replace only the input representation of an existing model.","The WER gains on CHiME-3 may partly reflect that ASR systems are trained on LogMel features, so enhancement in the same feature space avoids an extra linear-to-Mel conversion and its associated mismatch."],"forward_implications":["If Mel-McNet works as reported, multichannel speech enhancement can be run in the Mel domain with a roughly 60% reduction in FLOPs and no loss in perceived quality or ASR accuracy.","The enhanced LogMel output can be consumed directly by an ASR model, so the enhancement front-end and recognition backend can share the same feature representation.","The CPU real-time factor of about 0.799 (0.755 for the network, 0.044 for the vocoder) suggests the framework is deployable in real time without a GPU.","Mel-scale compression appears to preserve spatial information well enough for the narrow-band and full-band spatial modules of McNet to still separate target speech from noise."],"supporting_citations":[{"why":"Supplies the four-module McNet backbone that Mel-McNet modifies and the linear-frequency baseline whose FLOPs are compared.","marker":"[8]"},{"why":"Defines the handcrafted triangular Mel filterbanks used for STFT-to-Mel compression and for computing the Mel-PRM target.","marker":"[17]"},{"why":"Provides the Vocos neural vocoder that converts enhanced LogMel spectra to waveforms for evaluation.","marker":"[18]"},{"why":"Earlier Mel-spectrogram enhancement framework that motivates direct Mel-domain processing and the vocoder reconstruction setup.","marker":"[12]"},{"why":"Provides the CHiME-3 dataset and the official evaluation and test sets used for all enhancement and ASR comparisons.","marker":"[20]"},{"why":"One of the online multichannel baselines compared on quality, complexity, and word error rate.","marker":"[15]"},{"why":"Another online streaming baseline that Mel-McNet is compared against on CHiME-3.","marker":"[7]"},{"why":"DNSMOS P.835 is the non-intrusive perceptual metric used to measure speech quality in the main comparison.","marker":"[30]"}],"fun_headline_variants":["Mel-scale processing matches linear baseline at 60% lower cost","Mel-frequency domain cuts enhancement compute 60%","Mel-McNet: same speech quality, 60% cheaper computation","Mel-scale makes multichannel enhancement 60% lighter"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that comparing Mel-McNet, whose enhanced LogMel must be turned into a waveform by a neural vocoder, against baselines that produce waveforms directly is a fair test of Mel-scale processing; if the vocoder itself is responsible for the observed DNSMOS and WER behavior, the claimed benefit of Mel-scale processing is not cleanly isolated.","fun_headline_variants_meta":{"raw":{"variants":["Mel-scale processing matches linear baseline at 60% lower cost","Mel-frequency domain cuts enhancement compute 60%","Mel-McNet: same speech quality, 60% cheaper computation","Mel-scale makes multichannel enhancement 60% lighter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000345,"raw_usage":{"total_tokens":1867,"prompt_tokens":890,"completion_tokens":977,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":908}},"tokens_in":506,"tokens_out":977,"duration_ms":8161,"temperature":1.0,"reasoning_tokens":908,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:11:27.482524+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the McNet backbone through the same pipeline as Mel-McNet: convert McNet's enhanced STFT output to LogMel, reconstruct with the same Vocos vocoder, and evaluate DNSMOS and WER on CHiME-3. If this McNet plus Vocos baseline matches Mel-McNet's numbers, the 60% complexity reduction would be confirmed but the quality and ASR claim would not be due to Mel-scale processing.","supporting_citations":[{"cited_title":"Blstm supported gev beamformer front-end for the 3rd chime challenge,","cited_arxiv_id":null,"evidence_quote":"Supplies the four-module McNet backbone that Mel-McNet modifies and the linear-frequency baseline whose FLOPs are compared."},{"cited_title":"Deep neural mel-subband beamformer for in-car speech separation,","cited_arxiv_id":null,"evidence_quote":"Provides the Vocos neural vocoder that converts enhanced LogMel spectra to waveforms for evaluation."},{"cited_title":"Multichannel long-term streaming neural speech enhancement for static and moving speakers,","cited_arxiv_id":null,"evidence_quote":"Earlier Mel-spectrogram enhancement framework that motivates direct Mel-domain processing and the vocoder reconstruction setup."},{"cited_title":"The third ‘chime’ speech separation and recognition challenge: Dataset, task and baselines,","cited_arxiv_id":null,"evidence_quote":"Provides the CHiME-3 dataset and the official evaluation and test sets used for all enhancement and ASR comparisons."},{"cited_title":"Embedding and beamform- ing: All-neural causal beamformer for multichannel speech en- hancement,","cited_arxiv_id":null,"evidence_quote":"One of the online multichannel baselines compared on quality, complexity, and word error rate."},{"cited_title":"Unsupervised speech enhancement based on mul- tichannel nmf-informed beamforming for noise-robust automatic speech recognition,","cited_arxiv_id":null,"evidence_quote":"Another online streaming baseline that Mel-McNet is compared against on CHiME-3."}],"review_version":1}