acceptodds
Under review as a conference paper at ICLR 2027

WHICH PARTIALS MOVE: A 12-TET PULL ABOVE THE BAND EDGE OF A NEURAL CODEC DECODER

Abstract

Neural synthesis pulls output tuning toward 12-TET, as recently reported for melto-audio vocoders. We ask where inside a neural audio codec that happens. At low bitrate a decoder transmits a tone’s lower partials and invents the rest above a band edge that rises with bitrate within a codec. The transmitted fundamental stays put through EnCodec, whereas the vocoder we measure moves it by 3.1 cents (a cent is a hundredth of a semitone). The invented partials land most of the way onto the harmonic ladder of the nearest 12-TET pitch, and a pitch estimator fitting across the edge averages the transmitted and invented partials, reporting a displacement no partial underwent: a median of 4.37 cents over twenty-four stimuli. Where the invented partials land follows the decoder’s fine-tuning corpus. When only EnCodec’s decoder is fine-tuned, so that every arm reads bit-identical tokens, retuning its corpus by 33 cents moves the decoder’s grid by 32.7 ± 2.8 cents at seed-level spread, while grid-preserving resamplings of comparable or larger size leave it in place. Classical codecs show no pull. At a 440 Hz reference six of eleven measurable neural conditions show the effect, four of them by more than five cents. Those four hold under every variance treatment and span two codec families; the other two fall below their own detection floors. The practical exposure falls on full-band harmonic fitting, ours more than salience-based tools. After one round trip, 22 of 200 Carnatic and 11 of 60 Turkish makam recordings switch to fitting 12-TET better than their own corpus’s measured degrees, while a period tracker is largely spared. The remedy is simple where it applies: fit no partial above the edge. The band-edge account replaced a pre-registered one that failed, and its registered pass on fifteen fresh checkpoints tests a three-tier prediction, giving p = 0.070 against always predicting the middle tier and 0.375 against always predicting the most common tier.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
neural audio codecs, audio tokenizers, tuning bias, 12-TET, band edge, partial regeneration, pitch estimation, measurement artefacts, music information retrieval, Carnatic music, Turkish makam, fine-tuning corpus, causal intervention, pre-registration

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_BSqBOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how neural audio codecs affect pitch structure and its measurement. Controlled synthetic stimuli and per-partial spectral analysis show that, in the tested EnCodec settings, the fundamental and lower partials are largely preserved while some upper partials shift toward the harmonic ladder of the nearest 12-TET fundamental. An estimator combining these partials can consequently report an intermediate pitch displacement. The authors investigate the source through decoder-only corpus interventions that keep test tokens fixed, examine consequences for tuning analysis of real recordings, and propose restricting pitch estimation to partials below the measured band edge.

Critical strengths

1. The paper provides useful mechanistic insight by locating changes in individual partials. Figure 1 and Table 1 distinguish an essentially preserved fundamental, two spectral lines at an edge partial, and displaced upper partials. This explains why an aggregate pitch estimate can differ from the displacement of any individual partial. The contribution therefore goes beyond reporting overall pitch error and identifies a concrete measurement issue when analyzing audio processed by a neural codec.

2. The decoder corpus intervention is convincing. The encoder and quantizer are frozen, so each fixed test input produces identical tokens across decoder arms; only the decoder’s fine-tuning corpus changes. Shifting the corpus by 33 cents moves the preferred grid by approximately 32.7 ± 2.8 cents, while the fundamental remains essentially unchanged and grid-preserving resampling controls do not produce a corresponding shift (Section 3.3). The reported spread is across seeds. This intervention, together with directionally consistent results in other decoders, supports an effect of the corpus on upper-partial placement rather than merely showing that fine-tuning reduces reconstruction error.

Critical weaknesses

1. The magnitude of the impact in commonly used pitch-analysis pipelines is not yet sufficiently quantified. The paper includes real-music experiments, but its main error magnitudes and mitigation results come from the authors’ full-band harmonic estimator. With pYIN, the proportion of Carnatic recordings whose preferred fit switches to 12-TET falls from 11% to 3%. For librosa, the main result is a count of recordings moving toward 12-TET, without comparable displacement magnitudes or a distortion-matched control. The practical impact therefore depends strongly on the analysis tool. More specific evidence is needed to identify which common workflows incur consequential errors and how large those errors are (Section 3.4; Appendix F; Table A31).

2. Mitigation is primarily validated for measurement accuracy, while improvements to reconstructed-audio fidelity remain limited. Fitting only partials below the band edge reduces pitch-estimation error but does not restore displaced upper partials, and direct validation is primarily limited to the authors’ estimator. Decoder fine-tuning provides a possible route to improvement, but has cross-corpus tradeoffs: fine-tuning on Carnatic music reduces scale-degree error on that corpus while increasing it on makam recordings. For applications requiring faithful preservation of input tuning, the paper should more clearly distinguish measures that improve analysis reliability from those that bring reconstructed partials closer to the input, including their conditions and costs (Sections 3.3 and 3.4; Appendices E and F).

Questions for authors
  1. How large is the impact in commonly used pitch analysis, and which settings require intervention? Using the existing real recordings and the already tested librosa pipeline, please report changes in estimated tuning in cents with uncertainty, and include a traditional-codec control with distortion comparable to EnCodec. Together with the existing pYIN results, please clarify which common workflows are more or less affected and when the changes are sufficient to alter practical analysis conclusions.
  2. What effective measures are currently available when faithful preservation of input tuning is the goal? Please distinguish reducing pitch-estimation error from bringing the decoded upper partials themselves closer to the input. For the former, could band-limiting or an equivalent procedure be directly evaluated in one of the already tested off-the-shelf tools, with its requirements for edge estimation and usable low-frequency information specified? For the latter, please use the existing bitrate and fine-tuning experiments to explain which settings reduce upper-partial displacement, the associated reconstruction-quality or computational costs, and whether the improvement transfers to other musical corpora. This would clarify how far the existing remedies go.
Reviewer_Fh3sOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies why neural audio codecs shift musical pitch toward Western 12-TET tuning and where inside the codec this happens. At low bitrate, the decoder transmits the lower harmonics of a tone and invents the harmonics above a cutoff frequency, called the band edge. The fundamental and the transmitted harmonics stay in place, while the invented harmonics land near the harmonics of the nearest 12-TET pitch. A pitch estimator that fits all harmonics averages the two groups and reports a pitch shift that no harmonic actually underwent. The paper is mainly an analysis paper and does not propose a new model or method. The overall findings are that this pull toward 12-TET appears in some neural codecs, mostly EnCodec, and not in classical codecs, that it depends on the tuning of the decoder’s training music and does not require the quantizer, and that it is much smaller on real music and on speech than on synthetic tones. The authors also show that it can bias tuning analysis of non-Western music, and that estimating pitch only from harmonics below the band edge reduces it.

Critical strengths
  • The paper shows at the level of individual harmonics that the fundamental does not move while the invented harmonics do, which prior work measuring whole recordings could not separate.
  • The authors report their failed pre-registered hypothesis and collect all limitations in one table, which makes the paper transparent.
Critical weaknesses
  • The paper is very hard to read and follow, I think most of the paragraphs are AI generated. The authors disclose this and they mentioned that they verified the content, but even if the content is verified you should at least make some edits for clarity.
  • I think the contribution is incremental. Berendes et al. (2025, 2026) already showed that neural audio synthesis pulls tuning toward 12-TET, that the training corpus is the likely source, and that model capacity affects the size of the effect. This paper adds a finer localization of the same effect, and the proposed fix of fitting only below the band edge is simple once the mechanism is known.
  • The effect is narrow. Four of the six pulling conditions are EnCodec, and two of the six pull by less than a cent, below their own detection floors (Table A29). The authors themselves conclude that the mechanism is a property of particular decoders.
Questions for authors
  • Why was EnCodec chosen for most experiments, and do the real-music and downstream results hold for other codecs that pull, such as WavTokenizer? What design choices make some decoders place invented harmonics on the grid while others, such as DAC 44k, do not?
  • What are the practical implications for models built on codec tokens, such as music generation models, and should codec designers change how they train decoders?
Reviewer_vdxAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies tuning bias in neural audio codecs using controlled harmonic stimuli, reference detuning, and measurements of individual partials. It finds that, in some codec conditions, the fundamental remains approximately unchanged while regenerated upper partials move toward the harmonic ladder associated with a nearby 12-TET pitch. The authors describe this behavior using a band edge and a ladder fraction, investigate its relationship to the training corpus through decoder-only fine-tuning, and examine its consequences for tuning analysis on Carnatic and Turkish makam recordings.

Critical strengths

The analysis of individual partials provides a more detailed explanation of the phenomenon. The paper distinguishes an actual displacement of the fundamental from a change in the output of a harmonic pitch estimator. This usefully refines recording-level tuning measurements and identifies a concrete way in which codec distortion can affect pitch measurement.

The decoder-only corpus interventions provide reasonably convincing evidence. These experiments support the claim that training data can change the decoder's tuning preferences. In particular, the approximately matched shift between corpus retuning and attractor displacement, together with grid-preserving resampling controls, provides stronger support than a correlation between training statistics and output bias.

Critical weaknesses

The overall contribution remains limited in scope. Berendes et al. (2025, 2026) have already studied tuning bias and its mitigation through fine-tuning. This paper adds meaningful partial-level analysis and corpus interventions, but its strongest evidence concerns specific spectral behavior under controlled conditions. The broader practical significance remains insufficiently established.

Practical significance beyond the custom estimator is only partly characterized. The authors’ full-band fit is particularly sensitive to regenerated harmonics, while pYIN is less affected. The librosa experiment supports a directional change but does not establish its magnitude or isolate it with a distortion-matched control. Although the filtering checks support robustness, applicability to less harmonic or more strongly polyphonic material remains uncertain.

Task-level degradation attributable specifically to tuning bias is not yet well quantified. The Opus controls support the directional findings, but much of the improvement from band limitation reflects generic codec damage. The expanded makam experiment provides a descriptive accuracy decrease, yet corresponding controls were not repeated on the same recordings and reference pool. This limits attribution of the additional task-level loss to tuning bias.

Questions for authors

Magnitude and attribution in the standard-tool experiment: For the reported librosa experiment, can the authors quantify the distribution of within-recording tuning-offset changes and compare it with a distortion-matched control? The directional count is informative, but does not establish how large the effect is or how much is specific to the learned tuning bias. This analysis would affect my assessment of practical significance.

Comparable controls for the identification task: Using the existing 235 makam recordings, can the authors report the Opus-control and band-limited results with the same reference pool and evaluation protocol as the EnCodec arm? Such descriptive paired comparisons would help determine whether the observed accuracy loss exceeds generic codec damage and is reduced by the proposed mitigation. They would strengthen the task-level evidence without treating the expanded dataset as an independent confirmatory replication.

Sensitivity to the regeneration criterion: Could the level-only sensitivity analysis also be reported for the 15 preregistered checkpoints, keeping the original prediction rule and tier boundaries fixed? This would clarify whether the predictive conclusions remain robust when regeneration is identified without the spectral-placement condition.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.