{"id":"d0fda85a-e095-408d-bec6-8b894f5c3254","arxiv_id":"2507.00229","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"CTFT-Net jointly reconstructs magnitude and phase for speech bandwidth extension and reports lower log-spectral distance than NU-Wave, WSRGlow, NVSR, and AERO, but its own tables and core equation contain inconsistencies.","lead":"CTFT-Net is a speech super-resolution network that works on complex time-frequency spectrograms, aiming to restore both the missing high frequencies and the phase of low-quality speech. The authors report better spectral-distance scores than several prior systems on the VCTK dataset for upsampling from 2 kHz to 48 kHz.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (1) mis-defines complex convolution with a minus sign in the imaginary path, so the core architecture is either wrongly implemented or wrongly described; the central complex-domain claim cannot be evaluated until this is resolved.","rationale":"We focus on Eq. (1) because it is the narrowest gate through which every other claim must pass. The paper's stated novelty is a complex-domain U-Net with global attention; if the fundamental complex convolution is misdefined, the 'complex' part of the architecture is vacuous or mis-specified, and the reported performance cannot be attributed to the described method. A reader cannot reproduce the system from the text, because Eq. (1) does not define complex convolution. This is independent of any dataset-specific result: even if the numbers in Table 1 were exactly as reported, the method itself would be undefined. The AP-BWE row in Table 1 (LSD 1.016 vs 1.06 at 2 kHz) is a separate concern about the scope of the SOTA comparison, but it is secondary: the abstract's named baselines (NU-Wave, WSRGlow, NVSR, AERO) do not include AP-BWE, so the claim as worded is not directly falsified by that row. The impossible NISQA-MOS values in Table 3 (e.g., 11.47) also raise evaluation reliability questions, but they appear only in the ablation and are more plausibly typos than a systematic error affecting LSD. The equation error, by contrast, is an unambiguous mathematical inconsistency in the core method. We therefore agree with the reader that this is the weakest assumption. A code check or reimplementation is the single decisive test: it determines whether the paper contains a typo or whether the network actually computes a nonstandard transform, and in either case the submitted text needs a correction before the claimed architecture can be accepted.","tokens_in":10667,"tokens_out":10208,"duration_ms":102150,"concrete_test":"Inspect the released source code (or re-implement Eq. (1) from the paper) and run a minimal forward pass on a known complex input and kernel. Compare the imaginary output to Wr*Si_in + Wi*Sr_in. If the implementation uses the plus sign, Eq. (1) is a typo and the architecture is standard complex convolution; if it uses the minus sign, the operation is nonstandard and the complex-domain architecture claim fails. Reference: Section 2.2, Eq. (1).","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 2.2, Eq. (1) defines the 2D complex convolution that underpins all encoders, decoders, skip blocks, the CGAB, and the complex conformer. The real-part update Wr*Sr_in - Wi*Si_in matches standard complex multiplication (Wr + jWi)(Sr_in + jSi_in). The imaginary-part update, however, is written as Wr*Si_in - Wi*Sr_in. The standard imaginary part is Wr*Si_in + Wi*Sr_in. A minus sign here changes the operation from complex convolution to a nonstandard bilinear map; it is not complex conjugation (which would give Wr*Si_in - Wi*Sr_in only if the real part were Wr*Sr_in + Wi*Si_in). Thus Eq. (1) as printed is not a valid complex convolution. Since every downstream block is defined through this equation, the submitted text either describes a network that is not complex-valued or contains an error in the load-bearing formula. No code is provided, so the reader cannot determine which case holds; if the implementation follows Eq. (1), the claimed 'complex domain' processing is not what is claimed, and if it does not, the paper's core formula is wrong. The central claim that CTFT-Net is a high-fidelity complex-domain network therefore rests on an unresolved premise at the entry point of the architecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CTFT-Net, a complex-valued U-Net for speech super-resolution that takes a complex STFT input, processes it through complex encoders/decoders, complex skip blocks, a complex global attention block (CGAB), and a complex conformer bottleneck, and is trained with a combination of time-domain SI-SDR loss and multiresolution STFT loss. The authors report experiments on the VCTK dataset with target 48 kHz and 16 kHz sampling rates and claim that CTFT-Net outperforms NU-Wave, WSRGlow, NVSR, and AERO, particularly for extreme 2 kHz to 48 kHz upsampling, while jointly recovering magnitude and phase without a vocoder.","tokens_in":10962,"tokens_out":3109,"duration_ms":35047,"significance":"If the claims were substantiated, the paper would make a useful engineering contribution: a single complex-domain network that reconstructs both magnitude and phase for bandwidth extension, with an attention mechanism spanning time and frequency axes, would be attractive compared with vocoder-based pipelines. The ablation study is also a good-faith attempt to isolate the effects of CGAB placement, loss choices, and bottleneck architecture. However, the empirical evidence as presented is not reliable: the central performance claim is contradicted by the paper's own table, the core complex-convolution definition is wrong as written, and the ablation table contains impossible metric values. There is no released code, no checkpoints, and no uncertainty quantification, so the claimed improvements cannot be independently verified. The strengths of the paper are conceptual and architectural, not empirical at this stage.","major_comments":[{"comment":"The central claim that CTFT-Net achieves the lowest LSD among all baselines is contradicted by the numbers in Table 1. For the 2 kHz input, the proposed method reports LSD 1.06, while AP-BWE (also in the table) reports 1.016; for the 4 kHz input, the proposed method reports 0.96, while AP-BWE reports 0.92. The text states that \"for each LR input, CTFT-Net achieves the lowest LSD compared to all baselines,\" which is not true if AP-BWE is counted as a baseline. The comparison must be restated to name only the baselines for which the claim actually holds, or the table must be corrected if the AP-BWE numbers are in error.","section":"4.1, Table 1"},{"comment":"Equation (1) does not define a valid complex convolution. For a complex kernel W = Wr + jWi and input Sin = Sr + jSi, the standard complex multiplication gives Si0 = Wr * Si_in + Wi * Sr_in, whereas the paper writes Si0 = Wr * Si_in - Wi * Sr_in. This is not a typo in sign convention: it changes the operation from complex convolution to a nonstandard bilinear map. Since the paper explicitly says that every complex layer, including CGAB and the complex conformer, implements Eq. (1), the submitted text either describes a network that is not performing complex convolution or contains a formula error in the load-bearing definition. This must be resolved, ideally by providing code or a corrected derivation.","section":"2.2, Eq. (1)"},{"comment":"Table 3 contains NISQA-MOS values that are outside the valid range of the metric. NISQA-MOS is a mean opinion score on a 1-5 scale, yet rows P7, P8, and P9 report values of 11.47, 11.19, and 11.47, respectively. These values are physically impossible, which indicates a data-entry error or an undefined modification of the metric. Because Table 3 is the basis for the ablation conclusions, all rows must be re-verified and recomputed.","section":"4.2, Table 3"},{"comment":"Row P10, which removes the SI-SDR loss, reports an LSD of 0.88 — better than the proposed CTFT-Net configurations in P6.1 (1.06) and P6.2 (1.01). The paper claims that combining SI-SDR loss with MR-STFT loss improves audio quality, and the ablation is used to support that claim, but P10 shows the opposite for LSD. The authors must explain this discrepancy, for example by discussing loss weighting or the trade-off between LSD and SI-SDR, or by correcting the table.","section":"4.2, Table 3, row P10"},{"comment":"No code, checkpoints, or evaluation scripts are provided, and no error bars, confidence intervals, or per-utterance statistics are reported for any metric. The paper states \"Source code of the model will be available after acceptance,\" which prevents a reviewer or reader from verifying the central empirical claims, especially given the inconsistencies in Tables 1 and 3. Without these artifacts or statistical measures, the claimed superiority over SOTA systems is not reproducible.","section":"3.3, 4.1"}],"minor_comments":[{"comment":"Equation (2) has a formatting issue: the summation symbols and parentheses are not properly typeset, making it hard to see that the spectral convergence and log-magnitude losses are summed over resolutions.","section":"2.6, Eq. (2)"},{"comment":"Row label P6 is used in the text but the table contains P6.1 and P6.2, making the reference ambiguous; the text should say which configuration is meant.","section":"Table 3"},{"comment":"The conclusion repeats the claim of \"strong performance\" without acknowledging the discrepancies in Table 1 or the impossibility of some numbers in Table 3; a more cautious summary would be appropriate.","section":"5, Conclusion"}],"recommendation":"reject","confidential_remarks":"The paper has several deep issues that cannot be fixed by local copyediting. Table 1 contradicts the headline claim, Eq. (1) misdefines the core operation, Table 3 contains impossible NISQA-MOS values, and the SI-SDR ablation undercuts the stated loss contribution. Given that the empirical claims are the main contribution, I recommend rejection. If the authors can correct the formula, re-run and re-report all experiments, and provide code, a fresh submission would be worth considering."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the architecture: a fully complex U-Net for speech super-resolution that puts global attention on both time and frequency axes (CGAB) and inserts a complex conformer in the bottleneck. That specific combination doesn't appear in the prior work I know, and the motivation—joint magnitude and phase reconstruction without a vocoder or post-processing—is reasonable. The paper also runs a fairly thorough ablation and reports large LSD improvements for extreme upsampling (2 kHz to 48 kHz). If those numbers are right, that's practically useful.\n\nThe soft spots are serious, though. The equation for complex convolution, Eq. (1), has the wrong sign in the imaginary part. Standard complex convolution gives Wr*Si_in + Wi*Sr_in; the paper prints a minus. That's not a typo you can shrug off: the whole network, including CGAB and the conformer, is built on this operation. Without code, a referee cannot tell whether the implementation matches the text or not. Either way, the paper's description of its own load-bearing component is wrong.\n\nTable 1 directly contradicts the headline claim. AP-BWE beats CTFT-Net at 2 kHz and 4 kHz input, yet the text says CTFT-Net achieves the lowest LSD against all baselines. That's a factual inconsistency in the paper's main result. Table 3 has NISQA-MOS values over 11, which are outside the plausible range for a MOS predictor—likely a unit or labeling error. And the ablation line P10 (without SI-SDR) improves LSD to 0.88 from 1.06, undermining the claim that adding SI-SDR helps. None of these are explained.\n\nSo the paper is not ready as is. The central claim does not survive contact with its own data, and the core formula is unverifiable. That said, the architecture is worth thinking about, and the errors look fixable. I'd send this to a serious referee—not because it's near acceptance, but because the underlying idea is interesting and the mistakes are identifiable. A good reviewer could force the equation to be corrected, the baselines to be fair, and the numbers to be sane. I wouldn't cite it in its current form.\n\nIf the authors fix the equation, release code and checkpoints, and straighten out the tables, this could become a useful engineering contribution. As submitted, it's a salvageable idea wrapped in a flawed write-up.","headline":"A clever architecture undermined by a wrong core equation and inconsistent tables; the idea has merit, but the paper as submitted doesn't hold together.","tokens_in":11517,"tokens_out":3004,"would_cite":false,"duration_ms":34224,"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 fully complex U-Net with global time-frequency attention claims to reconstruct speech magnitude and phase jointly, beating four state-of-the-art bandwidth-extension models on VCTK, with its largest wins at 2-to-48 kHz upsampling.","keywords":["speech super-resolution","bandwidth extension","complex-valued neural network","complex global attention","phase reconstruction","multi-resolution STFT loss","SI-SDR loss","VCTK"],"falsifier":"Inspect the released training code (the paper says source will be available after acceptance) for the complex-convolution layer. If the imaginary branch computes $W_r \\ast S_i - W_i \\ast S_r$ as written, retraining with the standard $W_r \\ast S_i + W_i \\ast S_r$ and comparing LSD on VCTK 2-to-48 kHz upsampling would settle whether the claimed improvement depends on the non-standard transform; if instead the code uses the standard update, then Eq. (1) misstates the architecture and the paper's description of its core mechanism is wrong.","tokens_in":10472,"feed_emoji":"🎙️","tokens_out":8131,"duration_ms":86628,"temperature":0.7,"pith_summary":"CTFT-Net is a speech bandwidth-extension model that works entirely on complex-valued time-frequency spectrograms, so it predicts phase as well as magnitude instead of discarding phase or relying on a vocoder. The paper claims that on the VCTK corpus it beats four published super-resolution systems—NU-Wave, WSRGlow, NVSR, and AERO—on log-spectral distance, with the largest gains at extreme upsampling from 2 kHz to 48 kHz. The mechanism is a complex global attention block that models correlations along both the time and frequency axes, a complex conformer in the bottleneck, and a loss that combines time-domain SI-SDR with a multi-resolution real-part STFT loss. A careful reader would care because high-frequency content and phase drive perceived speech quality, and the result suggests a single complex U-Net can do 24x bandwidth extension without post-processing tricks.","feed_headline":"Complex U-Net beats four baselines at 24x speech upsampling","feed_subtitle":"It reconstructs phase as well as magnitude and reaches a log-spectral distance of 0.62 at 2-to-48 kHz upsampling.","key_machinery":"The load-bearing object is the complex-valued 2D convolution of Eq. (1), which the paper uses to build encoders, decoders, skip blocks, attention blocks, and the conformer; it lets the network treat each spectrogram as a complex tensor rather than handling magnitude and phase separately. Around that operation, the complex global attention block (CGAB) reshapes encoder features along the time and frequency axes, applies separate dense layers to each axis, and recombines them to model inter-phoneme and harmonic correlations. A complex conformer in the bottleneck adds multi-head self-attention and local convolution, and the loss combines real-part multi-resolution STFT terms with SI-SDR so the model is supervised in both frequency and time domains.","core_discovery":"The central claim is that magnitude and phase of a low-resolution speech spectrogram can be regenerated together by a U-Net whose every layer is complex-valued. The network takes an STFT of the low-resolution waveform, processes it through eight complex encoders and eight complex decoders with complex skip connections, inserts a complex global attention block after the first and seventh encoders to capture inter-phoneme and inter-frequency correlations, and places a complex conformer in the bottleneck to mix local and global context. It is trained by summing a real multi-resolution STFT loss (spectral convergence plus log-magnitude on the real part, at three STFT resolutions) with a time-domain SI-SDR loss. On VCTK the paper reports the lowest LSD among the four compared systems at 2, 4, 8, and 12 kHz input to a 48 kHz target, for example 1.06 versus 1.10 for NVSR at 2 kHz and 0.62 versus 0.87 at 12 kHz, while leaving SI-SDR essentially flat, which the authors read as absence of added noise. The paper also claims this removes the need for a vocoder, for copying input low-frequency bands at the output, and for flipping or reusing the input phase.","pith_inferences":["If Eq. (1)'s imaginary update is what was actually implemented, the network is not performing standard complex convolution; retraining with the standard update $W_r \\ast S_i + W_i \\ast S_r$ would test whether the reported gains persist, because the paper does not clarify whether the equation or the code is authoritative.","The CGAB's dual-axis global attention on complex spectrograms is a natural component to reuse in speech enhancement, dereverberation, or text-to-speech systems that need high-frequency and phase fidelity, although the paper does not test those tasks.","The paper only evaluates fixed input resolutions on speech from VCTK; applying the same loss and attention design to variable bandwidths or to music would show whether the mechanism generalizes beyond the reported setting."],"forward_implications":["A vocoder-free, post-processing-free bandwidth-extension pipeline that predicts phase directly could replace vocoder-dependent systems such as NVSR.","Extreme upsampling ratios, from telephony-grade 2 kHz to 48 kHz, become feasible with a single network, with the reported LSD dropping from 1.06 at 2 kHz input to 0.62 at 12 kHz input.","Joint training with multi-resolution STFT and SI-SDR improves perceived quality, as measured by PESQ, while keeping intelligibility and SI-SDR nearly unchanged, indicating no noisy artifacts at the band edge.","The complex global attention block, which attends to both axes of a complex spectrogram, is the component that the ablation study credits for the main gains over frequency-only attention."],"supporting_citations":[{"why":"NU-Wave is one of the four baseline super-resolution systems CTFT-Net is compared against.","marker":"[13]"},{"why":"WSRGlow is a glow-based waveform generative baseline that CTFT-Net must beat.","marker":"[14]"},{"why":"NVSR is the strongest compared baseline and the source of the post-processing and vocoder shortcuts CTFT-Net claims to avoid.","marker":"[15]"},{"why":"AERO is a spectral-domain audio super-resolution baseline used in the comparison.","marker":"[16]"},{"why":"The VCTK multi-speaker corpus supplies all training and test speech for the reported comparisons.","marker":"[30]"},{"why":"Defines SI-SDR, which CTFT-Net uses both as a training loss and as an evaluation metric.","marker":"[19]"},{"why":"Provides the spectral convergence and log-magnitude STFT losses that make up the real multi-resolution STFT loss.","marker":"[20]"},{"why":"Provides the conformer architecture whose complex-valued version is placed in the CTFT-Net bottleneck.","marker":"[29]"},{"why":"Supplies the frequency-transformation block used as the ablation baseline that CGAB is meant to beat.","marker":"[12]"}],"fun_headline_variants":["Complex U-Net beats four baselines at 24x upsampling","Phase-aware complex U-Net leads speech super-resolution","CTFT-Net: complex attention for 24x speech upsampling","Reconstructing phase improves 24x speech super-resolution","Complex global attention lifts extreme speech upsampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Eq. (1) as printed, with the imaginary update described as real kernel times imaginary input minus imaginary kernel times real input, is what the network actually computes; if it is a typo, the paper misstates its central building block, and if it is not, the network applies a non-standard transform.","fun_headline_variants_meta":{"raw":{"variants":["Complex U-Net beats four baselines at 24x upsampling","Phase-aware complex U-Net leads speech super-resolution","CTFT-Net: complex attention for 24x speech upsampling","Reconstructing phase improves 24x speech super-resolution","Complex global attention lifts extreme speech upsampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000767,"raw_usage":{"total_tokens":3416,"prompt_tokens":978,"completion_tokens":2438,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":2352}},"tokens_in":594,"tokens_out":2438,"duration_ms":20592,"temperature":1.0,"reasoning_tokens":2352,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:20:59.670880+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the released training code (the paper says source will be available after acceptance) for the complex-convolution layer. If the imaginary branch computes $W_r \\ast S_i - W_i \\ast S_r$ as written, retraining with the standard $W_r \\ast S_i + W_i \\ast S_r$ and comparing LSD on VCTK 2-to-48 kHz upsampling would settle whether the claimed improvement depends on the non-standard transform; if instead the code uses the standard update, then Eq. (1) misstates the architecture and the paper's description of its core mechanism is wrong.","supporting_citations":[{"cited_title":"Bandwidth extension of musical au- dio signals with no side information using dilated convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"NU-Wave is one of the four baseline super-resolution systems CTFT-Net is compared against."},{"cited_title":"Speech band- width extension using generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"WSRGlow is a glow-based waveform generative baseline that CTFT-Net must beat."},{"cited_title":"NU-GAN: High resolution neural upsampling with GAN","cited_arxiv_id":"2010.11362","evidence_quote":"NVSR is the strongest compared baseline and the source of the post-processing and vocoder shortcuts CTFT-Net claims to avoid."},{"cited_title":"Speech super resolution gen- erative adversarial network,","cited_arxiv_id":null,"evidence_quote":"AERO is a spectral-domain audio super-resolution baseline used in the comparison."},{"cited_title":"Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (ver- sion 0.92),","cited_arxiv_id":null,"evidence_quote":"The VCTK multi-speaker corpus supplies all training and test speech for the reported comparisons."},{"cited_title":"Nu-wave: A diffusion probabilistic model for neural audio upsampling,","cited_arxiv_id":null,"evidence_quote":"Defines SI-SDR, which CTFT-Net uses both as a training loss and as an evaluation metric."},{"cited_title":"Wsrglow: A glow-based waveform generative model for audio super-resolution,","cited_arxiv_id":null,"evidence_quote":"Provides the spectral convergence and log-magnitude STFT losses that make up the real multi-resolution STFT loss."},{"cited_title":"A time-frequency attention module for neural speech en- hancement,","cited_arxiv_id":null,"evidence_quote":"Provides the conformer architecture whose complex-valued version is placed in the CTFT-Net bottleneck."},{"cited_title":"Phasen: A phase-and- harmonics-aware speech enhancement network,","cited_arxiv_id":null,"evidence_quote":"Supplies the frequency-transformation block used as the ablation baseline that CGAB is meant to beat."}],"review_version":1}