{"id":"0b0614b4-7838-4e9b-9990-b05e1668017f","arxiv_id":"2607.16657","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A training-only change to residual vector quantization orders codebook stages from bass to treble, yielding better audio quality and more predictable bitrate scaling with identical inference.","lead":"HARP is a new training recipe for neural audio codecs that teaches each compression stage to specialize on a different frequency range, from bass to treble, without changing how the codec runs at deployment. It reports better reconstruction scores and listening ratings than standard RVQ codecs, with the largest gains at low bitrates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (12) does not guarantee band-local group contributions; Table 3 shows the low-mid group collapses onto bass, so the central 'frequency-ordered partition' claim is only partially supported.","rationale":"The reader's weakest_assumption is exactly Eq. (12)'s locality assumption, which I independently identify as the most load-bearing concern. This is where the method's theoretical justification and empirical evidence diverge: the paper claims frequency-ordered partitioning, but Table 3 shows that groups 0 and 1 collapse to the same center. This is not merely a cosmetic issue; it challenges the causal mechanism proposed for HARP's improvements. If the partition is not realized, the explanation that 'each group refines its target band' and that harmonic coherence arises from cumulative decoding is undermined. The alternative concern about group dropout confounding the DAC comparison is mitigated by the paper's statement (Section 5.3) that the DAC baseline uses identical hyperparameters as HARP except no band supervision, implying the same group dropout. Thus the core unresolved issue is the band-locality of Eq. (12). The paper's own Table 3 and discussion acknowledge the low-band collapse but do not analyze its consequences. The proposed concrete test would settle whether the band supervision actually produces band-limited contributions and whether the method's benefit is attributable to the claimed mechanism. Since this is a serious gap in validating the central claim but not a demonstrated fatal flaw, the existing CONDITIONAL verdict remains appropriate.","tokens_in":14139,"tokens_out":5203,"duration_ms":54994,"concrete_test":"Compute the spectral energy concentration of x_k on a held-out test set (e.g., MUSDB18-HQ): for each group k, measure the fraction of x_k's total spectral energy that falls within its target band (low 0–1 kHz, low-mid 1–4 kHz, etc.). If group 1's x_k has substantial energy below 1 kHz (e.g., >30% of its energy), the band supervision is not isolating the intended band, confirming the collapse seen in Table 3. Additionally, retrain with non-learnable, fixed band centers (µ_k frozen at initial values) and compare performance; if HARP's advantage over DAC persists unchanged, then the learned-band collapse is not harmful, but if the advantage disappears, the method's benefit depends on the un-intended low-frequency concentration rather than the claimed harmonic-aware partition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of HARP is that Eq. (12) isolates group k's spectral contribution x_k = G(sg[z_prev] + z_k) - sg[x_prev], so applying the mel-weighted band loss (Eq. 15) steers group k toward its target band. This relies on the subtraction yielding a band-local signal, but G is a nonlinear convolutional decoder; the subtraction does not prevent x_k from containing energy outside the intended band. Because Eq. (15) compares x_k to the full-band waveform x (not a band-filtered residual), group k can be rewarded for reconstructing lower-frequency content when the decoder spreads x_k across the spectrum. Table 3 provides direct evidence of this failure: learned centers for groups 0 and 1 both converge to 0.08 (~228 Hz) instead of the initialized 0–1 kHz / 1–4 kHz split. The low-mid group did not specialize; it consolidated with the bass group. Thus the 'frequency-ordered partition' and 'each group refines its target band' claims are only partially realized. The paper presents no quantitative analysis of x_k's band energy concentration, and the visual mel-spectrogram claim is not sufficient. If x_k is not band-local, the observed improvements over DAC may stem from a different mechanism (e.g., simply forcing early stages to focus on low frequencies) rather than the proposed harmonic-aware per-band refinement. This weakens the central claim as stated in the abstract and section 6.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HARP, a training-only modification of RVQ-based neural audio codecs in which the nine RVQ stages are partitioned into four ordered groups (3-2-2-2 allocation) associated with low-to-high frequency bands. For each group, the decoded waveform increment is computed via Eq. (12) and supervised with a learnable mel-band-weighted L1 loss (Eq. 15); cumulative decoding retains lower-stage latents as context. At inference the architecture is identical to standard RVQ. Experiments on music, speech, and general audio compare HARP against DAC and a parameter-matched BSCodec, reporting SI-SDR, KAD, PESQ, STOI, MUSHRA, synthetic harmonic-coherence tests, and ablations. The central claim is that HARP outperforms both standard RVQ and parallel band decomposition while preserving harmonic coherence and requiring no architectural change.","tokens_in":14535,"tokens_out":4299,"duration_ms":45042,"significance":"If the central claim holds, HARP is a useful and economical contribution: it imposes a frequency hierarchy on RVQ entirely through the training loss, leaves inference and the token stream unchanged, and offers more graceful bitrate truncation than standard RVQ. The paper also includes a synthetic harmonic-coherence experiment that directly targets the proposed mechanism, an ablation isolating the main design choices, and public code/models. The main risk is that the spectral-locality assumption underpinning Eq. (12) is not demonstrated and is partially contradicted by Table 3, and the headline objective and subjective claims are reported without variance or significance testing. With additional diagnostics and more careful statistical reporting, the contribution would be solid.","major_comments":[{"comment":"The central mechanism treats x_k = G(sg[z_prev] + z_k) - sg[x_prev] as group k's isolated spectral contribution. Because G is a nonlinear convolutional decoder, the subtraction does not guarantee band-local content; energy from x_k can spread across the spectrum. Eq. (15) compares x_k to the full-band waveform x, not to a band-filtered residual, so a group can be rewarded for reconstructing lower-frequency content even if its intended band is not respected. Table 3 gives direct evidence that the intended frequency ordering only partially emerges: the learned centers for groups 0 and 1 both converge to 0.08 (~228 Hz), so the low-mid group consolidated with the bass group instead of splitting at the initialized 0-1 kHz / 1-4 kHz boundary. The paper needs a quantitative band-energy-concentration diagnostic for each x_k (e.g., fraction of energy in the target band) and either revised claims","section":"§4.4, §4.5, §6.2 (Eq. 12, Eq. 15, Table 3)"},{"comment":"The headline comparisons are point estimates without variance, confidence intervals, or significance tests. Table 2 reports SI-SDR and KAD on a single test set per domain; a +0.25 dB music SI-SDR gain over DAC may be within run-to-run or test-set noise. Table 5 spans bitrates but has no error bars, and Table 7 reports a PESQ gain of +0.09 while stating STOI is 'not significant' without showing any test. Table 6's MUSHRA medians/IQRs are not accompanied by an inferential statistic, yet the text claims HARP scores 'significantly above' DAC. The authors should provide standard errors, per-seed or per-item variability, and appropriate paired significance tests for the main objective and subjective claims.","section":"§6.1, §6.4, §6.5 (Tables 2, 5, 7, 6)"},{"comment":"The MUSHRA protocol is incompletely reported. A MUSHRA test normally includes an explicit low anchor; no anchor scores are given, and the conditions listed omit BSCodec entirely, so the subjective comparison cannot assess the parallel-decomposition baseline. The claim of perceptual improvement rests on a single trained model, 12 participants, and median/IQR summaries. The authors should report the anchor condition, participant screening details, per-item scores, and a paired test (e.g., Wilcoxon signed-rank) for the DAC-HARP difference.","section":"§6.5 (Table 6)"}],"minor_comments":[{"comment":"Notation is inconsistent: Eq. (12) defines \\hat{x}_k, but Eq. (15) and later text use x_k; the paper should use one symbol consistently.","section":"§4.4-4.5"},{"comment":"Table 3 compares HARP's learned centers \\mu_k with DAC's empirical mel centroids from 'tier-diff contributions' on n=50 test items, with no variance for either column. These are different quantities; please clarify what is being compared and report dispersion.","section":"§6.2 (Table 3)"},{"comment":"The text says band losses are computed 'regardless of n', but Algorithm 1 computes them only for active groups k=0,...,n-1. Clarify whether inactive groups receive band losses when group dropout is active.","section":"§4.7, Algorithm 1"},{"comment":"The BSCodec baseline is described as 'scaled down' to match parameter count, but no details are given on quantizer capacity, bitrate matching, or training hyperparameters. A short specification would strengthen the fairness argument.","section":"§5.3"},{"comment":"STOI is described as a 'direct indicator of speech intelligibility' but is computed here on reconstructed speech at 7.7 kbps; the reader would benefit from a sentence noting that STOI saturates at high quality, which likely explains the small difference.","section":"§5.4"}],"recommendation":"major_revision","confidential_remarks":"The central idea is attractive and the paper is generally well organized, but the main risk is the spectral-locality assumption in Eq. (12). I would ask the authors to provide direct diagnostics of band-energy concentration for each group contribution, and to either relax the 'frequency-ordered partition' claim or support it with evidence. The missing variance/significance analysis in Tables 2, 5, 6, and 7 is also important because the headline gains are modest in some domains. This is fixable within the manuscript's scope, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. HARP is a new training-only trick for RVQ codecs: group the stages, decode cumulatively, and supervise each group's waveform increment with a mel-weighted loss. That's genuinely not in the cited literature—MUFFIN does latent FFT allocation, BSCodec splits bands architecturally, and [25] uses two branches. The paper also does a lot right: it gives a clear mechanism, runs ablations that show cumulative decoding is the main driver, includes MUSHRA with a hidden reference, and is honest about one of its own limitations (Table 3 shows the low-mid group collapsed onto bass). Inference cost is unchanged, which is a real practical advantage.\n\nThe soft spots are real but not fatal. The core assumption in Eq. (12)—that subtracting the previous group's output isolates a band-local signal—is never verified. The decoder is nonlinear, so x_k can contain energy outside its target band, and the loss compares x_k to the full-band waveform, not a band-filtered residual. The paper's own Table 3 shows groups 0 and 1 both converge to ~228 Hz, so the 'frequency-ordered partition' is only partially realized. The authors acknowledge this in the text but still repeat the strong claim in the abstract and conclusion; that needs to be toned down or backed with a diagnostic of x_k's band energy concentration. Second, the main metric tables have no error bars or significance tests, and the MUSHRA has 12 participants. Third, BSCodec is missing from the listening test with a procedural reason, which weakens the perceptual comparison. Fourth, the DAC comparison is confounded by group dropout: HARP uses it, DAC doesn't, and there's no ablation separating that effect.\n\nNone of this kills the paper. The cumulative-decoding mechanism looks sound—removing it costs 1.4 dB, and the random-phase harmonic test shows HARP keeps phase coherence near DAC and far above BSCodec. The circularity burden is low: main results are on held-out sets and listening test. The method is likely useful for low-bitrate codecs even if the exact band split isn't as clean as advertised.\n\nI'd send this to peer review. A careful referee can push for the missing diagnostics, but the core idea deserves engagement. I'd bring it to the reading group and consider citing it as related work, though I wouldn't build on it until the band-locality question is settled.","headline":"Training-only spectral hierarchy for RVQ is a real idea; the evidence is suggestive but the band-locality assumption and missing controls make the strongest claims premature.","tokens_in":15020,"tokens_out":4844,"would_cite":true,"duration_ms":43379,"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":"HARP trains neural audio codecs to put bass in early stages and treble in later ones, improving quality with no extra inference cost.","keywords":["neural audio codec","residual vector quantization","spectral entanglement","frequency hierarchy","harmonic coherence","subband supervision","training-only method","bitrate scalability"],"falsifier":"Compute the actual spectrum of x_k for each group on held-out audio; if group 1's waveform increment retains substantial energy below 1 kHz beyond the floor weight, or if removing the band loss changes the learned centers and the claimed gains vanish, the isolation assumption fails. A simpler check: if the learned centers of groups 0 and 1 remain identical at 0.08 after longer training with a stronger band weight, the frequency-ordered hierarchy is not being achieved.","tokens_in":14021,"feed_emoji":"🎧","tokens_out":4063,"duration_ms":35090,"temperature":0.7,"pith_summary":"This paper tries to show that the residual vector quantization stages of a neural audio codec can be made spectrally specialized by changing only the training loss, not the network architecture. Standard RVQ codebooks capture an unpredictable mix of frequencies, so truncating stages for low bitrates removes bass or treble erratically. HARP partitions the nine RVQ stages into four frequency-ordered groups, decodes each group's contribution cumulatively over all lower groups, and supervises each contribution with a soft, learnable mel-band weight. The paper reports that this yields better objective and listening-test scores than both standard RVQ and a parallel band-split codec, with largest gains at low bitrate, and that inference stays identical to standard RVQ.","feed_headline":"Training trick makes neural audio codecs drop treble before bass","feed_subtitle":"Changing only the loss gives codecs a frequency hierarchy, improving low-bitrate audio quality with zero inference cost.","key_machinery":"The load-bearing mechanism is the waveform increment defined in Eq. (12): x_k = G(sg[z_prev] + z_k) - sg[x_prev], which subtracts the stopped-gradient output of prior groups from a decoder pass that keeps the prior cumulative latent as context. This isolates, in the gradient path, the spectral contribution of group k while still giving the decoder access to lower-frequency content, so that mid and treble reconstruction is supervised in the presence of its harmonic foundation. Two auxiliary pieces support it: subband contribution supervision (the band loss applied to x_k rather than to the cumulative output) and soft band weighting, a learnable Gaussian over mel bins with a fixed floor, which","core_discovery":"On the paper's own terms, the central discovery is that spectral hierarchy can be imposed on RVQ through the loss only: by supervising each stage group's isolated waveform increment (Eq. 12) with a Gaussian-weighted mel loss, the learned codebooks organize themselves so that early groups carry low-frequency content and later groups add high-frequency detail, while the decoder always receives the cumulative lower-frequency latent as context. This preserves harmonic coherence that parallel band decomposition loses, because overtones are reconstructed with knowledge of their fundamentals. The paper claims consistent SI-SDR and kernel-audio-distance gains over the standard RVQ codec it builds on","pith_inferences":["The paper's Table 3 shows that groups 0 and 1 both converge to the same learned center (~228 Hz), so the five bass/low-mid codebooks are not splitting at the 1 kHz boundary as initialized. A likely extension is to test whether reallocating stage counts or increasing the band loss weight forces a cleaner split.","Because the band parameters are learned, the hierarchy adapts to data: on datasets with little low-frequency energy, groups may re-organize. An open extension is to measure how the learned centers shift across domains.","The stop-gradient in Eq. (12) means codebook entries are updated only through commitment/codebook losses, not through the band loss. A testable variant is to allow a small band-loss gradient into the codebook entries of the current group, which may improve band localization without breaking isolation."],"forward_implications":["If correct, any RVQ-based codec can gain frequency-prioritized bit allocation without architectural change, so low-bitrate truncation becomes predictable: bass survives, treble is dropped first.","The single unified token stream is preserved, unlike parallel band-split codecs, so HARP can be dropped into existing codec language-model pipelines unchanged.","The gap over standard RVQ grows at low bitrates (average +0.6 dB at 2.6 kbps vs +0.5 dB at 7.7 kbps), meaning the benefit is largest exactly where bitrate is scarce.","Harmonic coherence is measurably better than parallel decomposition: on aligned-phase synthetic tones HARP reaches phase coherence 0.988 vs the parallel codec's 0.914 with high variance, directly supporting the cross-band-context claim.","The method is orthogonal to the quantizer variant, so the same loss can be applied to other RVQ-style codecs."],"fun_headline_variants":["Loss-only tweak makes codecs prioritize bass then treble","New training loss orders codec stages by frequency for better audio","HARP: reorganize codec codebooks with a single loss change","Zero-cost codec upgrade: just change the loss function","Frequency-aware loss improves neural audio codecs with no inference cost"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that the subtraction in Eq. (12) really isolates group k's spectral contribution; because the decoder is a nonlinear convolutional network, the difference of two decoded waveforms is not guaranteed to be band-local, and the paper provides no direct diagnostic that x_k is confined to its target band.","fun_headline_variants_meta":{"raw":{"variants":["Loss-only tweak makes codecs prioritize bass then treble","New training loss orders codec stages by frequency for better audio","HARP: reorganize codec codebooks with a single loss change","Zero-cost codec upgrade: just change the loss function","Frequency-aware loss improves neural audio codecs with no inference cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000296,"raw_usage":{"total_tokens":1513,"prompt_tokens":663,"completion_tokens":850,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":407,"completion_tokens_details":{"reasoning_tokens":763}},"tokens_in":407,"tokens_out":850,"duration_ms":8108,"temperature":1.0,"reasoning_tokens":763,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T20:18:47.189687+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the actual spectrum of x_k for each group on held-out audio; if group 1's waveform increment retains substantial energy below 1 kHz beyond the floor weight, or if removing the band loss changes the learned centers and the claimed gains vanish, the isolation assumption fails. A simpler check: if the learned centers of groups 0 and 1 remain identical at 0.08 after longer training with a stronger band weight, the frequency-ordered hierarchy is not being achieved.","supporting_citations":[],"review_version":1}