{"id":"d02e3212-e222-4c5a-bff5-0018e3a731e6","arxiv_id":"2608.00572","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A single flow-matching model performs speech bandwidth extension across continuously varying cutoff frequencies by treating the observed low-band spectrum as an in-context prompt and infilling the masked high band.","lead":"AnyBand turns speech bandwidth extension into a fill-in-the-missing-frequencies problem, where a single model works for any input cutoff within a continuous range. It conditions high-frequency generation on the observed low-frequency spectrum, and beats prior fixed-bandwidth systems on spectral reconstruction while staying competitive in perceptual quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Clean-low-band assumption is structural: AnyBand copies the observed band (Eq. 10) and is trained/evaluated only on clean mel-masked inputs, so real noisy/codec-degraded inputs are outside the validated scope.","rationale":"The reader identified the clean-low-band premise as the weakest assumption; I agree and consider it the single most load-bearing concern. The paper's Eq. (10) copies the observed band into the final mel estimate, and training/validation masks clean 48 kHz mel spectrograms, so the model never sees a degraded low band. The in-context formulation further relies on the low band as a trustworthy prompt plus F0/UV conditioning extracted from that same band. This makes the method a pure BWE system whose practical reach is narrower than the abstract's 'practical scenarios' language. However, this is a boundary-condition limitation rather than a flaw in the core comparative claim: on clean band-limited inputs the reported experiments and ablations are coherent, the curriculum and adversarial components are internally consistent, and the improvements over baselines are large. The lack of error bars and code, plus the unspecified CFG unconditional model, lower confidence but do not change the verdict. A targeted degradation experiment would settle the scope. Hence the reader's CONDITIONAL verdict remains appropriate; I recommend UNCHANGED.","tokens_in":20855,"tokens_out":5806,"duration_ms":75705,"concrete_test":"Build a degraded low-band evaluation set: start from 48 kHz VCTK test utterances, apply realistic degradation to the low band (e.g., add speech-shaped noise at 20/10/0 dB SNR below the cutoff, or encode the band-limited signal with Opus at 16/24 kbps), then mask high frequencies as in the paper. Run the released AnyBand inference and baselines (AudioSR, FLowHigh) on both clean and degraded inputs. If AnyBand's LSD/HF-LSD and NISQA gaps over baselines shrink or reverse under degradation (especially at low SNR), the clean-low-band assumption is load-bearing and the 'practical scenarios' claim must be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that a single model produces better high-frequency reconstruction than fixed-bandwidth baselines across continuous cutoffs—is supported only for clean, mel-masked low bands. In Eq. (10), the final output is constructed as Mfb = (1-M)⊙fM + M⊙Mhat, i.e., the observed band is copied verbatim into the output. All training and evaluation inputs are constructed by masking clean 48 kHz mel spectrograms (Experiment Settings), so fM is always exactly the ground-truth low band. Under real telephony/codec degradation, the low band is noisy or spectrally distorted; any artifact in fM is propagated into the output with no denoising pathway, and the spectral prompt plus F0/UV conditioning (extracted from the same degraded input) provide corrupted cues to the generative model. This is not merely an 'untested' edge case—it is a structural premise of the in-context formulation: the method is a pure BWE, not a joint enhancement+extension system, and the abstract's appeal to 'practical scenarios' is broader than the validation. The concern does not refute the core comparative results on clean band-limited speech; it narrows the scope and justifies a conditional verdict.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AnyBand, a bandwidth-extension (BWE) framework that treats the observed low-frequency mel spectrum as a spectral prompt and reconstructs the missing high-frequency region via masked conditional flow matching. A single model is trained over continuously sampled cutoff frequencies using an Easy-to-Balanced curriculum, with a frequency-aware Diffusion Transformer encoder/decoder and three auxiliary discriminators for spectral realism, cross-band envelope coherence, and harmonic consistency. Experiments on VCTK and EARS report consistent LSD/HF-LSD improvements over NU-Wave 2, AudioSR, FLowHigh, and Fre-Painter at both standard and irregular cutoffs, with competitive NISQA, COL, and STOI scores and a small listening test.","tokens_in":21264,"tokens_out":7299,"duration_ms":97004,"significance":"If the claims hold, AnyBand is a useful step toward practical BWE: a single model handles a continuous range of cutoffs without cutoff-specific retraining, and the spectral-infilling formulation is simple and general. The paper's strengths include the continuous-cutoff training scheme, explicit frequency-axis modeling, a physically motivated discriminator set, out-of-domain evaluation, and honest reporting of ablation results where individual components produce mixed effects. The measured gains are empirical rather than derived, so the central contribution is the architecture/training recipe plus the comparative evaluation. The main risks are the clean-low-band assumption, an underspecified classifier-free guidance procedure, and the lack of an ablation isolating the spectral-prompt mechanism.","major_comments":[{"comment":"The final output copies the observed band (1-M)⊙fM verbatim, and all training/evaluation inputs are clean 48 kHz mel spectrograms with bins above the cutoff masked. The abstract's claim that the method addresses 'practical scenarios' with diverse cutoffs is therefore broader than the validated scope. On noisy or codec-degraded low bands (telephony, archival audio), AnyBand has no denoising pathway and would propagate low-band artifacts into the output; the spectral prompt and F0/UV conditioning (A.1) are also extracted from the same degraded input. Either add experiments with noisy/codec-degraded low-band inputs or restrict the claims to clean band-limited speech.","section":"Eq. (10) / Experiment Settings"},{"comment":"Classifier-free guidance is applied in velocity space, but no unconditional training scheme is described. Eq. (8) trains only the conditional velocity Gθ(Mt, fM, t, c); nothing in the architecture or losses specifies how v̂∅ is obtained (e.g., prompt dropout, mask dropout, or setting fM/c to a null token). Since CFG with scale 1.4 is used for all reported results, the inference procedure is not reproducible until this is specified.","section":"Eq. (12) / Training and Sampling"},{"comment":"The core 'in-context prompt' hypothesis is not ablated. The closest variants remove frequency modules or F0/UV, but no experiment replaces the observed-spectrum prompt with a scalar cutoff embedding or otherwise removes the prompt while keeping the mask. Without such an ablation, the paper does not demonstrate that conditioning on the low-band spectrum per se, rather than on the cutoff/mask, drives the continuous-cutoff capability.","section":"Table 4 / Component Analysis"}],"minor_comments":[{"comment":"The phrase 'may therefore may require retraining' contains a duplicated 'may'.","section":"Introduction"},{"comment":"LSD is defined as the RMS of log10-power differences, but the standard log-spectral distance is expressed in dB (10 log10). Please clarify the units or add the factor 10 so the reported absolute values are interpretable.","section":"Eq. (15)"},{"comment":"The protocol describes 5 input rates × 5 utterances × 5 systems plus ground-truth references, which gives 150 samples (or 125 without references), but the text states each listener rates 130 audio samples. Please reconcile the count.","section":"B.4 Subjective Evaluation"},{"comment":"At 4 kHz, removing D_cross improves LSD/HF-LSD over the full model. The text acknowledges isolated improvements, but the bolding/underlining in the table makes the pattern visually confusing; a footnote would help.","section":"Table 6"}],"recommendation":"major_revision","confidential_remarks":"The work is competent and the comparative results are mostly convincing on clean band-limited speech. The main issues are fixable: an overbroad practical-scope claim, an underspecified CFG training detail, and a missing prompt-ablation. I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: AnyBand is the first BWE system I've seen that genuinely treats variable cutoffs as one infilling problem rather than a set of fixed conversions, and it backs that up with careful ablations and consistent wins on spectral metrics. Worth a serious look. The stress-test concern is on the mark—the method is validated only for clean masked low bands, and Eq. (10) copies the observed band verbatim, so noisy/codec-degraded inputs are outside scope. That's a genuine limitation, but it doesn't undermine the comparative results on clean band-limited speech; it just means the abstract's 'practical scenarios' promise is broader than the evidence.\n\nWhat's new: the prompt-based spectral conditioning with an explicit mask over a continuously sampled cutoff range, plus the frequency-aware DiT and the three complementary discriminators. The package is new even if each ingredient exists. The flow matching math is clean: the linear path gives v = M - ε, the masked loss is properly normalized, and the endpoint-focused adversarial rollout is a sensible way to avoid backprop through the whole ODE. The ablations are the strongest part—they isolate the curriculum, the frequency modules, F0/UV conditioning, and each discriminator, and the irregular-cutoff results show the continuous training is doing real work.\n\nSoft spots, in order: (1) The clean-low-band assumption. Every experiment masks clean 48 kHz mel spectrograms. In telephony, archival audio, or hearing-aid inputs, the low band is noisy or codec-degraded. Because the observed band is copied into the output (Eq. 10) and the prompt conditions on that same degraded band, AnyBand will propagate artifacts. It's a pure BWE, not enhancement, and that should be stated as a scope condition rather than implied universal applicability. (2) Objective tables have no error bars or significance tests; the gains are consistent across settings, which helps, but I'd like confidence intervals. (3) Code isn't released, and the CFG unconditional model isn't described—minor, but relevant for reproducibility. (4) The subjective test is small (15 listeners, 5 utterances per condition) but that's typical for this subfield.\n\nThe citation pattern looks fine—self-cites are for auxiliary components (CodecFlow, PACE) and the ablations show removing them hurts, so no red flag.\n\nWho's it for: the speech-restoration and audio-super-resolution community. Anyone building deployment-ready BWE should read it, and it's a good benchmark candidate. I'd send it out; it deserves referee time. My own verdict: accept after revision, with the scope clarification as a required change.","headline":"Solid, well-ablated BWE paper whose real novelty is the continuous-cutoff infilling framing; the clean-observed-band assumption narrows applicability but doesn't break the core result.","tokens_in":21677,"tokens_out":2085,"would_cite":true,"duration_ms":24779,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single model can extend speech bandwidth at any cutoff by treating the observed low band as a spectral prompt and filling the missing high band.","keywords":["bandwidth extension","spectral infilling","flow matching","in-context conditioning","speech super-resolution","frequency-aware transformer","adversarial refinement","cutoff curriculum"],"falsifier":"Take a frozen AnyBand model and compare two test sets built from the same clean utterances: one masked with the paper's clean-cutoff protocol, and one whose low band is first passed through a telephone codec or corrupted with additive noise before masking. If high-band LSD and listening scores degrade sharply on the corrupted-prompt set, the spectral-prompt assumption fails. Separately, fine-tune a fixed-bandwidth baseline per cutoff (e.g., one model at 4 kHz, one at 8 kHz) and compare at those cutoffs; if the per-cutoff specialists clearly beat the single AnyBand, the unification claim loses","tokens_in":1804,"feed_emoji":"🎙️","tokens_out":4863,"duration_ms":114952,"temperature":0.7,"pith_summary":"AnyBand claims that speech bandwidth extension can be solved as a single in-context spectral-infilling task rather than as a collection of fixed source-to-target conversions. Given any cutoff frequency, the model treats the observed low-frequency spectrum as a prompt and generates the missing high band, with a frequency mask marking what to fill. The paper argues that this formulation lets one model handle a continuous range of input bandwidths, standard and irregular, without cutoff-specific retraining. Across in-domain VCTK and out-of-domain EARS data, AnyBand reports consistently lower spectral distortion than existing BWE systems while matching their perceptual quality, with the largest gains under severe band limitation. The practical point is that real speech arrives with diverse and unknown cutoffs, so a single mask-conditioned model could replace a zoo of bandwidth-specific systems.","feed_headline":"One model restores any missing speech band from the low band alone","feed_subtitle":"AnyBand treats the low-frequency spectrum as a prompt, beating fixed-bandwidth systems at every cutoff.","key_machinery":"The central mechanism is the spectral prompt plus a frequency mask: the retained low-band spectrum is fed to the generator as conditioning, and the binary mask specifies which mel bins to reconstruct, making every cutoff the same continuation task. Around this core, frequency-aware encoder/decoder with frequency-axis attention, a temporal DiT backbone, missing-band conditional flow matching with velocity prediction, an Easy-to-Balanced cutoff curriculum, and endpoint-focused adversarial refinement with three discriminators (multi-scale spectral, cross-band envelope, F0-aligned harmonic) carry the argument.","core_discovery":"AnyBand's central claim is that recasting BWE as in-context spectral infilling unifies multi-bandwidth extension in one model: the observed low-frequency spectrum is the prompt, a binary mask (1 above the cutoff) marks the missing region, and a flow-matching model generates the high band conditioned on that prompt. Trained over continuously sampled 1-16 kHz cutoffs with a missing-band objective and an Easy-to-Balanced curriculum, a frequency-aware transformer (frequency-axis encoder and decoder around a temporal DiT) captures cross-frequency cues, while three discriminators refine realism. The evidence: lowest LSD and HF-LSD at every standard and irregular cutoff on VCTK and EARS, with compe","pith_inferences":["If the observed spectrum really conveys speaker, prosody, and envelope cues, the same prompt-plus-mask infilling formulation could extend to music or general audio super-resolution by swapping the spectral prompt and mask.","Because the low band is copied verbatim into the output, real noisy or codec-degraded inputs would propagate low-band artifacts into the reconstructed high band; testing on degraded low bands is the natural stress test not covered by the clean-masking protocol.","The mask-conditioned formulation could let one deployed model serve devices with different bandwidth budgets by choosing the cutoff at inference time, without retraining.","The explicit frequency-axis attention suggests the model may be learning harmonic continuation rather than pure interpolation; a direct probe would be to evaluate cutoffs outside the trained 1-16 kHz range to see whether the harmonic structure extrapolates or silently collapses."],"forward_implications":["A single model can produce 48 kHz full-band speech from any input cutoff in the 1-16 kHz training range, including irregular cutoffs never seen as discrete training conditions.","Spectral reconstruction (LSD and HF-LSD) is consistently better than the compared fixed-bandwidth and flexible BWE baselines, with perceptual and intelligibility metrics remaining competitive.","Out-of-domain generalization to EARS speakers holds, suggesting that the spectral prompt carries speaker, content, and prosody cues rather than memorized cutoff-specific patterns.","The gains are largest when the missing band is wide, which is exactly the regime where practical telephony and archival audio suffer most.","The Easy-to-Balanced curriculum specifically lifts low- and medium-cutoff performance relative to both discrete and continuous uniform cutoff sampling."],"supporting_citations":[{"why":"Supplies the temporal DiT backbone and flow-matching speech-generation conditioning that AnyBand adapts with frequency modules.","marker":"Chen et al. 2025"},{"why":"Establishes acoustic-context infilling for speech, the principle AnyBand moves from the time axis to the frequency axis.","marker":"Le et al. 2023"},{"why":"Provides the conditional flow matching objective and velocity parameterization used for training the generative model.","marker":"Lipman et al. 2023"},{"why":"Supplies the endpoint-focused adversarial flow-matching refinement strategy AnyBand adopts for discriminator training.","marker":"Lee, Choi, and Lee 2024"},{"why":"Serves as a fixed-bandwidth baseline and supplies the LF/HF-LSD spectral evaluation protocol.","marker":"Han and Lee 2022"},{"why":"Serves as a versatile audio super-resolution baseline that AnyBand must beat on both standard and irregular cutoffs.","marker":"Liu et al. 2024"},{"why":"Serves as the Fre-Painter baseline, representing masked-representation learning for audio super-resolution.","marker":"Kim et al. 2024"},{"why":"Supplies the multi-scale spectrogram discriminator design underlying AnyBand's spectral-realism discriminator.","marker":"Kong, Kim, and Bae 2020"},{"why":"Supplies the Vocos vocoder that converts AnyBand's generated mel spectrograms into 48 kHz waveforms for evaluation.","marker":"Siuzdak 2024"}],"fun_headline_variants":["AnyBand: one model, any bandwidth, from low band alone","Single model restores any missing speech band using low band as prompt","Unified BWE: low band prompts spectral infilling across all cutoffs","One model handles every speech cutoff via in-context spectral infilling","AnyBand: low-frequency prompt fills any missing high band"],"cache_read_input_tokens":23424,"weakest_assumption_plain":"AnyBand assumes the observed low-frequency band is clean and trustworthy: the output copies that band verbatim, and all experiments create band-limited inputs by masking clean 48 kHz speech, so noisy or codec-degraded low bands are not validated.","fun_headline_variants_meta":{"raw":{"variants":["AnyBand: one model, any bandwidth, from low band alone","Single model restores any missing speech band using low band as prompt","Unified BWE: low band prompts spectral infilling across all cutoffs","One model handles every speech cutoff via in-context spectral infilling","AnyBand: low-frequency prompt fills any missing high band"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000771,"raw_usage":{"total_tokens":3268,"prompt_tokens":776,"completion_tokens":2492,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":2401}},"tokens_in":520,"tokens_out":2492,"duration_ms":19834,"temperature":1.0,"reasoning_tokens":2401,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:31:24.630802+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a frozen AnyBand model and compare two test sets built from the same clean utterances: one masked with the paper's clean-cutoff protocol, and one whose low band is first passed through a telephone codec or corrupted with additive noise before masking. If high-band LSD and listening scores degrade sharply on the corrupted-prompt set, the spectral-prompt assumption fails. Separately, fine-tune a fixed-bandwidth baseline per cutoff (e.g., one model at 4 kHz, one at 8 kHz) and compare at those cutoffs; if the per-cutoff specialists clearly beat the single AnyBand, the unification claim loses","supporting_citations":[{"cited_title":"23rd Annual Conference of the International Speech Communication Association, Interspeech 2022 , pages =","cited_arxiv_id":null,"evidence_quote":"Serves as a fixed-bandwidth baseline and supplies the LF/HF-LSD spectral evaluation protocol."}],"review_version":1}