{"id":"c5a43c69-f819-4009-8614-30ba77f8ac16","arxiv_id":"2607.24323","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"Vocos's 1D ConvNeXt backbone is effective for magnitude but fails at phase and phase-difference prediction; switching to 2D convolutions largely fixes the latter subtask.","lead":"Vocos, an efficient time-frequency neural vocoder, still lags BigVGAN on bandlimited mel inputs because its 1D-convolution backbone models magnitude well but phase poorly. The finding points future vocoder work toward architectures with stronger time-frequency inductive bias rather than more loss terms or scaling.","discovery_kind":"extension","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The Conv1D-vs-Conv2D comparison in Table 3 changes four things at once (conv dimensionality, channel widths 512/1536→32/64, block count 8→6, parameters 15M→37K), so the causal attribution of the failure to \"1D convolutions\" specifically is under-controlled.","rationale":"The reader identified the external-validity concern (phase-difference prediction from GT magnitude as a proxy for end-to-end vocoding failures). I partially agree: that concern is real, but the paper itself hedges appropriately in §5.3, and the §5.2 oracle ablation (Vocos-Phase performing worst even with GT magnitude) provides independent evidence that phase modeling is the bottleneck in the actual vocoding setting, which mitigates the proxy-validity worry. I instead locate the soft spot one level earlier, inside the diagnostic experiment itself: the Conv1D→Conv2D comparison that carries the causal claim is confounded on at least three axes simultaneously. This is a correctness-risk concern about internal attribution, not a consensus disagreement. The direction of the claim is plausible — the 400× parameter reduction with comparable FLOPs and a large metric improvement is genuinely suggestive, and the Conv2D result beating the Masuyama benchmark is a real, useful finding on its own. But \"1D layers are the bottleneck\" is doing prescriptive work in the conclusion (guiding future architecture research), so the confound is load-bearing rather than cosmetic. I recommend CONDITIONAL rather than REJECT because: the empirical observations in Tables 2 and 3 are not in dispute, the oracle-magnitude/phase ablation is clean, and the proposed control experiment is cheap and would very likely confirm the claim — the paper needs one more ablation row, not a redesign. If the capacity-matched controls come back as predicted, this returns to ACCEPT unchanged.","tokens_in":9312,"tokens_out":2087,"duration_ms":73165,"concrete_test":"Run two controls on the Table 3 task: (1) a Conv2D variant with channel widths scaled so parameter count matches ~15M (same blocks/depth as Vocos where feasible); (2) a Conv1D variant with capacity reduced toward the Conv2D regime (e.g., 64–128 channels) and/or with the input projection replaced by a per-frequency-bin embedding that preserves frequency identity before 1D convs. If the 15M Conv2D model still succeeds and the capacity-matched Conv1D model still fails with L_wa near 0.6, the attribution to conv dimensionality is confirmed. If the small Conv1D model improves substantially, or the frequency-preserving-input Conv1D model succeeds, the bottleneck is capacity/input representation, not 1D convolutions.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim is that 1D convolutional layers are the bottleneck for phase-difference prediction, based on Table 3: Vocos backbone (Conv1D) achieves L_wa 0.646 / LSC −20.24 dB, while the Conv2D variant achieves 0.112 / −29.56 dB. But per §3.2, the Conv2D variant is not a dimensionality-only swap: it uses 6 blocks instead of 8, channel widths (32, 64) instead of (512, 1536), and ~400× fewer parameters. The comparison therefore cannot isolate \"1D vs 2D convolution\" as the causal factor. Three alternative explanations remain live: (a) the 15M-parameter Conv1D model is over-parameterized for this task and poorly optimized at 400k steps (the training regime in Table 1 was carried over from a different-scale benchmark model), and a capacity-matched Conv1D model might do much better; (b) the Conv1D model's input layer, which projects 513 frequency channels to 512 feature channels, destroys frequency-bin identity before any convolution sees it — i.e., the failure may be due to the channel-mixing input projection rather than the 1D temporal convolutions per se; (c) some interaction of depth/width with the wrapping-aware loss. Note the paper's own framing (§5.3) is careful (\"inductive biases ... in particular the use of 1D convolutions\"), but the headline finding as stated in the abstract and the reader's strongest_claim (\"identify 1D convolutional layers are hindering their accurate prediction\") asserts more than the ablation design supports. This matters because the paper's prescriptive conclusion — future work should build time-frequency inductive biases into the backbone — depends on the diagnosis being about conv dimensionality rather than, say, input representation or optimization scale. The rest of the paper (vocoder comparison, oracle ablations in §5.2) is well-controlled and unaffected by this concern.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript revisits the Vocos time-frequency vocoder to explain its residual quality gap relative to time-domain vocoders such as BigVGAN. The authors retrain both systems on LibriTTS with a shared, modern GAN recipe and a parameter-matched BigVGANv2-base, and confirm via objective metrics and a 16-listener MUSHRA test that the gap persists after controlling for loss functions, discriminators, and model size. Oracle ablations (Vocos-Mag / Vocos-Phase, Table 2) show the Vocos backbone predicts magnitude compatible with ground-truth phase almost perfectly, but cannot predict phase compatible with ground-truth magnitude — and that joint prediction beats phase-only prediction, implying Vocos's 'inconsistent' magnitude actively compensates its phase errors. Finally, the authors adapt the Vocos backbone to predict baseband/frequency phase differences (a non-autoregressive precursor of phase reconstruction) and find the Conv1D backbone fails (L_wa 0.646, LSC −20.24 dB) while a small Conv2D variant succeeds (0.112, −29.56 dB) and beats the Masuyama et al. benchmark, leading to the conclusion that 1D convolutions are the bottleneck for modeling time-frequency structure.","tokens_in":9781,"tokens_out":3120,"duration_ms":105547,"significance":"If the results hold, this is a useful diagnostic contribution to an actively debated question (why time-frequency vocoders lag time-domain ones). Particular strengths worth naming: matched retraining of both vocoders on the same data and recipe, a parameter-matched BigVGANv2-base control, the identification and correction of a concrete bug-level issue (the magnitude clamping threshold, Fig. 2), oracle ablations that yield a genuinely new and falsifiable insight (that Vocos's inconsistent magnitude prediction compensates phase error, implying STFT-consistency regularizers may be counterproductive), a MUSHRA test with corrected significance testing, and publicly posted stimuli. The phase-difference experiment is a clean idea and the 37K-parameter Conv2D model beating a 247K benchmark at ~6× fewer FLOPs is notable. The main limitation is that the causal attribution of the phase-diff failure specifically to 1D convolutions is not isolated by the current ablation design.","major_comments":[{"comment":"The Conv1D-vs-Conv2D comparison that supports the paper's headline conclusion changes four things simultaneously: convolution dimensionality, channel widths (512/1536 → 32/64), block count (8 → 6), and parameter count (15.0M → 37.1K, a ~400× reduction). The causal attribution to '1D convolutional layers' (abstract) is therefore under-controlled. Live alternatives include (a) the 15M Conv1D backbone being mis-scaled or under-optimized for this task at 400k steps with the vocoding-derived training recipe of Table 1, and (b) depth/width interactions with the wrapping-aware loss. A capacity-matched Conv1D control (e.g., a ~37K–250K parameter Conv1D model with comparable receptive field) is a cheap experiment and is needed before the abstract's identification claim is warranted. Absent that control, the claim should be moderated to the level of the more careful §5.3 wording ('inductive biases","section":"§3.2, Table 3"},{"comment":"An alternative explanation for the Conv1D backbone's failure is the input layer itself: the first Conv1D (k=7) projects 513 log-magnitude channels to 512 feature channels, destroying frequency-bin identity before any convolution operates. The Conv2D variant and the Masuyama benchmark both preserve frequency as a spatial axis end-to-end. The failure may therefore lie in the channel-mixing projection rather than in 1D temporal convolutions per se — a distinction that matters for the paper's prescriptive conclusion about future architectures. A control retaining 1D temporal convolutions but with a frequency-preserving front end (or a Conv2D model with a fully mixing front end) would separate these hypotheses.","section":"§3.2, Fig. 1"},{"comment":"The argument that failure at non-autoregressive phase-difference prediction from ground-truth log-magnitude explains Vocos's end-to-end phase errors rests on an unstated bridge assumption: the two settings differ in input (513-bin magnitude vs 80-bin mel), supervision (wrapping-aware L1 vs adversarial + mel losses), and task (predict differences vs predict wrapped phase jointly with magnitude). It is plausible, and §3 motivates it via the signal model, but the manuscript should state it explicitly as an assumption and ideally provide one corroborating observation — e.g., whether the Conv2D phase-difference model's LSC improvement (−29.56 dB) transfers when its integrated phase is paired with Vocos-predicted magnitudes, or at minimum a discussion of why L_wa/LSC on the subtask is expected to predict LSC in Table 2.","section":"§3, §5.3"}],"minor_comments":[{"comment":"SCOREQ values for the oracle rows (Vocos-Mag: 0.073/0.147; Vocos-Phase: 0.495/0.614) sit outside the conventional MOS-like range; please state the scale/normalization and confirm the values, since Vocos-Mag is described as outperforming 'almost any other experiment' partly on this metric.","section":"Table 2"},{"comment":"Phase-difference models are trained for 400k steps versus 2M for vocoders (Table 1). Please justify the budget or show convergence curves, since under-training the Conv1D model is one of the confounds in the Table 3 comparison.","section":"Table 1, §5.3"},{"comment":"The MUSHRA test uses 16 listeners rating 14 of 20 German utterances; please report listener screening (e.g., reference-identification criterion), confidence intervals in addition to the pairwise significance markers, and whether the female/male split in Fig. 4 shows any systematic condition-by-gender interaction.","section":"§5.1, Fig. 4"},{"comment":"LSC is defined (Eq. 5) as a phase-reconstruction metric given ground-truth magnitude, but in Table 2 it is computed from the vocoders' own output waveforms (where both magnitude and phase are predicted). A sentence clarifying this dual usage would help, since Vocos-Mag's −29.18 dB is not directly comparable to the end-to-end rows.","section":"Eq. (5), Table 2"},{"comment":"The magnitude-clamping observation (threshold 100 clipping up to 41% of training samples) is a nice practical finding; please state explicitly whether the 400 threshold is used in all retrained Vocos rows of Table 2, including 'Vocos (official†)'.","section":"§2.2, Fig. 2"},{"comment":"Eq. (3): state the bin-index convention (m = 0 or 1) for the linear-phase term 2πmH/N. Also check reference [30] (SSW 2019 for a vocoder-comparison paper with these authors seems early) and the line-broken URL of the accompanying website.","section":"Eq. (3), References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as a workshop-format (IWAENC-style) contribution: compact, one diagnostic idea, limited ablation budget. The science is honest and the hedged passages in §5.3 suggest the authors are aware of the control gaps in Table 3, but the abstract currently makes the strong causal claim. If the journal's scope accommodates short diagnostic/negative-result-style analyses, this is a good fit once the Table 3 attribution is tightened; the requested capacity-matched control is a small training run, so the revision burden is modest."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core here is the controlled diagnosis of why Vocos still trails BigVGAN on band-limited mels. They retrain both under matched losses and discriminators, run oracle Mag/Phase splits (Table 2), and show that letting magnitude be inconsistent actually helps the joint model compensate for phase errors. That finding is clean, new relative to the WaveNeXt/WaveHax/Siuzdak discussion, and immediately useful if you care about spectrogram-inversion architectures.\n\nThe MUSHRA (16 listeners, significance tests) and the magnitude-clamp fix are also done properly; the clamp histogram alone is a small but real service to anyone re-implementing Vocos. Circularity is low: external baselines, independent metrics, no fitted constants recycled as predictions.\n\nThe soft spot is real but localized. Table 3 attributes the phase-difference failure to “1D convolutions,” yet the Conv2D variant simultaneously drops to 6 blocks, (32,64) channels, and ~400× fewer parameters. So the causal arrow is under-controlled; over-parameterization, the initial frequency-mixing projection, or optimization scale remain live alternatives. The paper’s own prose in §5.3 is more careful than the abstract, but the headline claim still overreaches. That weakens the prescriptive close (“build better TF inductive biases”) without touching the earlier oracle and vocoder-gap results.\n\nWho it is for: anyone working on neural vocoders or phase reconstruction who wants a clear negative result on pure 1D backbones. Not a new model, not a theory paper. I would bring it to reading group for the ablation design and the Mag/Phase insight. It deserves peer review; a referee can simply ask for a capacity-matched 1D control. I would cite the oracle split and the matched BigVGAN gap.","headline":"Solid diagnostic ablations on Vocos; the 1D-vs-2D claim is overstated because Table 3 changes capacity and depth at once, but the rest of the evidence still holds.","tokens_in":10869,"tokens_out":468,"would_cite":true,"duration_ms":11170,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Vocos still trails time-domain vocoders because its 1D convolutions cannot capture the time-frequency structure required for accurate phase reconstruction.","keywords":["neural vocoders","phase reconstruction","Vocos","time-frequency","mel spectrogram","ConvNeXt","phasiness","phase differences"],"falsifier":"A Vocos-style model that retains 1D convolutions yet matches BigVGAN MUSHRA scores on the same band-limited mel inputs and training recipe, or a 2D-convolution Vocos that closes the listening-test gap while still accepting arbitrary conditioning representations.","tokens_in":10572,"feed_emoji":"🔊","tokens_out":891,"duration_ms":34677,"temperature":0.7,"pith_summary":"Time-frequency neural vocoders such as Vocos are attractive for efficiency, yet they still produce lower-quality speech than strong time-domain models. This paper shows the gap remains even after matching training losses, discriminators, and band-limited mel inputs. Ablations establish that the Vocos backbone models magnitude spectrograms easily but struggles with phase. When the same backbone is asked only to predict phase differences—the intermediate quantities classical reconstruction methods use—1D convolutions fail while a 2D variant succeeds with far fewer parameters. The authors conclude that future work must supply inductive biases that let the network see local time-frequency structure of speech without abandoning support for arbitrary input representations.","feed_headline":"Vocos fails at phase because of its 1D convolutions","feed_subtitle":"The same backbone with 2D layers fixes phase-difference prediction and isolates the real quality bottleneck.","key_machinery":"Phase-difference prediction (baseband and frequency phase differences) used as a non-autoregressive diagnostic task. Training the Vocos backbone solely on these differences, then integrating them by least squares, tests whether the architecture can capture the local time-frequency structure that classical phase-reconstruction methods exploit.","core_discovery":"The Vocos architecture is effective for magnitude modeling but substantially less so for phase. When the backbone is adapted to predict phase differences from ground-truth log-magnitude spectrograms, Conv1D layers perform poorly, whereas the identical design with Conv2D layers dramatically improves both wrapping-aware loss and log-spectral convergence. This isolates 1D convolutions as the component that prevents the network from modeling the time-frequency couplings needed for phase reconstruction.","pith_inferences":["The same 1D-versus-2D contrast may explain why other purely convolutional spectrogram-to-waveform models underperform relative to models that explicitly exploit time-frequency locality.","A hybrid that keeps a cheap 1D path for arbitrary conditioning while adding a lightweight 2D refinement stage could preserve Vocos efficiency while fixing phase.","The overlooked magnitude-clamping threshold suggests some published Vocos quality numbers may be limited by a hyperparameter rather than pure architecture.","If phase-difference prediction is a reliable diagnostic, future TF-vocoder papers should report classical consistency metrics alongside perceptual scores."],"forward_implications":["Time-frequency vocoders will not close the quality gap until their backbones better encode local time-frequency structure of speech.","Scaling the existing 1D Vocos architecture or simply adding magnitude/phase spectrogram losses is unlikely to remove the residual phasiness.","Allowing the network to predict inconsistent magnitudes currently compensates for phase errors; forcing consistency may hurt waveform quality.","Future designs need inductive biases for harmonic and local TF structure that do not rely on pitch priors or restrict the input representation.","Band-limited mel spectrograms remain a useful stress test that accentuates the domain gap between time-domain and time-frequency vocoders."],"fun_headline_variants":["Vocos 1D convolutions block accurate phase-difference prediction","Conv2D backbone fixes Vocos phase gaps that Conv1D cannot","Ablation shows Vocos handles magnitude well but fails on phase","1D layers hinder time-frequency couplings needed for phase","Swap to 2D convs and Vocos phase-difference loss drops sharply"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That success or failure at predicting phase differences from perfect magnitude spectrograms is a valid explanation for the phase errors that appear in the full end-to-end vocoder trained jointly on mel inputs with adversarial losses.","fun_headline_variants_meta":{"raw":{"variants":["Vocos 1D convolutions block accurate phase-difference prediction","Conv2D backbone fixes Vocos phase gaps that Conv1D cannot","Ablation shows Vocos handles magnitude well but fails on phase","1D layers hinder time-frequency couplings needed for phase","Swap to 2D convs and Vocos phase-difference loss drops sharply"]},"model":"grok-4.5","effort":"low","cost_usd":0.003102,"raw_usage":{"total_tokens":1044,"prompt_tokens":736,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":31024000,"prompt_tokens_details":{"text_tokens":736,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":233,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":736,"tokens_out":75,"duration_ms":4635,"temperature":1.0,"reasoning_tokens":233,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T18:05:46.562101+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"A Vocos-style model that retains 1D convolutions yet matches BigVGAN MUSHRA scores on the same band-limited mel inputs and training recipe, or a 2D-convolution Vocos that closes the listening-test gap while still accepting arbitrary conditioning representations.","supporting_citations":[],"review_version":1}