Hearing Meter Through the Group of Beat Positions: An Invariance Test for Rhythm-Blind Audio Metrics

Automatic scores for generated music, such as the Fréchet Audio Distance, compare the overall statistics of sound embeddings between real and generated clips. They can rate a drum loop as excellent even when its hits land at random, unmetered times that listeners hear as noise. We propose to treat the positions in a musical bar as a cyclic group: sixteenth notes in a 4/4 bar, for example, behave like integers that wrap around at 16. Within that structure we build rhythm perturbations whose effect on meter we know exactly. Some rearrangements provably keep a loop's metrical structure, some provably destroy it, and every one reuses the same drum hits, so timbre and the overall spectrum barely move. Running the same perturbations past human listeners and past each automatic metric shows which metrics are blind to meter. It also tests a specific hypothesis about what meter perception actually tracks.

The tools come from two areas of pure mathematics. Finite-field theory supplies the tests for when a map shuffles group elements bijectively and when it collapses them onto a subgroup. Partition theory supplies generating functions, which let us count exactly how surprising a loop's concentration on the beat grid is under a random-placement null. Neither tool has been used in audio evaluation.

Core Claim

Write a one-bar loop with $k$ onsets on an $N$-tick grid as the polynomial $O(x)=\sum_i x^{t_i}$ modulo $x^N-1$, where $t_i$ is the tick of the $i$-th hit. For a divisor $d$ of $N$, define $n_r$ as the number of onsets in residue class $r$ modulo $d$. A loop is metrical at level $d$ when its onsets pile into a few of these classes.

Two classes of rearrangement behave very differently. Multiplying every position by a unit $u$, meaning any $u$ with $\gcd(u,N)=1$, gives the map $t\mapsto ut$. It permutes the residue classes among themselves, so the multiset of class counts at every level, and the multiset of Fourier magnitudes at each character order, is exactly preserved. Multiplying by a non-unit is not a bijection: it merges hits onto a proper subgroup. A uniform reshuffle preserves only the hit count.

This is the same phenomenon that drives permutation-polynomial theory. A monomial $x^m$ permutes a finite field exactly when $\gcd(m,q-1)=1$. The proof in Permutation binomials over finite fields that a binomial $x^m+ax^n$ cannot permute $\mathbb{F}_p$ when $\gcd(m-n,p-1)\le\sqrt{p}-1$ rests on the image collapsing into cosets of a subgroup. The sharper exclusions of $\gcd(m-n,p-1)\in\{2,4\}$ in Nonexistence of Permutation Binomials of Certain Shapes come from Hermite's criterion, which detects collapse through power sums that should vanish. We prove no new binomial theorems. We borrow the lens: group characters detect coset concentration, unit maps preserve it, and non-unit maps collapse it.

The perceptual hypothesis this lens produces is concrete and falsifiable. Listener acceptability of a drum loop should be roughly invariant under unit dilations and translations, and should drop sharply under coset-breaking shuffles and collapsing maps. Embedding-distribution metrics should be roughly invariant under all of them.

The measurement side uses generating functions. The exact null probability that $k$ uniformly placed onsets have every class count at most $c$ at level $d$ is a coefficient extraction:

$$ P\big(\max_r n_r \le c\big)=\frac{[x^k]\Big(\sum_{j=0}^{c}\binom{N/d}{j}x^j\Big)^{d}}{\binom{N}{k}}. $$

This is the bounded-part counting familiar from restricted-partition work such as Biases in non-unitary partitions, which proves coefficient inequalities by manipulating generating functions directly. Combining these tail probabilities across the divisor lattice of $N$ gives a reference-free coset-concentration score.

Because the score uses maximum class occupancy over all residues rather than "on the downbeat," it is phase-invariant. Syncopated offbeat patterns still score as metrical. This is a deliberate guard against the failure reported in Improving Automatic Jazz Melody Generation by Transfer Learning Techniques, where intentional deviations were penalized as statistical outliers.

Research Program

The motivating evidence comes from four sources. A Benchmarking Initiative for Audio-domain Music Generation using the FreeSound Loop Dataset argues that distribution matching captures texture but not rhythmic grammar. ViSQOL v3 and A Unified Neural Architecture for Instrumental Audio Tasks show that sample-matching metrics fail in the other direction, penalizing plausible generative detail. SongEval and CMI-RewardBench document that learned judges and human raters entangle their rating dimensions; CMI-RM's musicality and alignment heads reach an SRCC of about 0.85. None of these works separates what a metric does under perturbations of known structure, which is what we build.

The work has three parts.

Perturbation suite. We start from the FreeSound Loop Dataset (2,608 loops) and the Looperman set (23,983 loops). Both are time-stretched to 120 BPM, so each clip is exactly one 2-second bar and needs no tempo estimation. We quantize detected onsets to an $N=48$ grid, which accommodates both sixteenths and eighth-note triplets. Each hit is resynthesized at its new position using a slice-and-crossfade renderer, producing these conditions:

  • an identity re-render that controls for slicing artifacts;
  • cyclic translations;
  • unit dilations $t\mapsto ut$ with $\gcd(u,48)=1$;
  • within-coset shuffles, which keep class counts at a coarse level but not at a finer one;
  • non-unit dilations, which collapse hits onto a subgroup;
  • uniform reshuffles with the hit count held fixed.

Scorer. We implement the generating-function score exactly in rational arithmetic, plus a differentiable version that replaces hard onsets with soft onset envelopes and class counts with roots-of-unity filters. The differentiable version can serve as a reranking or training signal.

Human calibration. We run a listening study modeled on the 140-participant Drumness and Preference protocol of Hung et al., with conditions balanced across loops and listeners.

Baselines

The do-nothing reference is the set of metrics the field already reports for this data: Fréchet Audio Distance and Inception Score computed exactly as in Hung et al. They report StyleGAN2 at FAD 2.22 and IS 6.08 on Looperman, and IS 5.24 on FSLD against 6.30 for real data. We compute these on each perturbation condition using the same embedding model and classifier. Beating them means our score separates conditions that humans separate, where FAD and IS do not. Concretely, our score's rank correlation with mean human Drumness across conditions should exceed theirs, with confidence intervals bootstrapped over loops and listeners.

The strongest learned competitors are two quality predictors. The first is a UTMOS-style predictor trained on SongEval, which the file reports at Pearson 0.912 with human ratings on full songs, against 0.61–0.66 for Audiobox-Aesthetic. The second is CMI-RM, reported at 78.20% accuracy on CMI-Pref musicality, compared with 65.80% for Gemini 3 Pro. Both are the strongest competitors because they are trained directly on human musicality judgments and could already encode meter. Neither was evaluated on short drum loops, and for rhythm-perturbation sensitivity specifically there is no reported result found on Lacuna. We run both off the shelf on identical clips, with no tuning on our listening data. As an extra comparison, we fine-tune each on the same held-out fraction of human ratings that we use to set our single combination weight across divisor levels, so the tuning budgets match.

A simple rhythm-feature baseline completes the comparison: inter-onset-interval entropy plus beat-tracker confidence, computed from the same quantized onsets.

Ablations isolate the key ingredients. Replacing the exact generating-function null with a Monte Carlo shuffle null tests whether exactness matters. Replacing the phase-invariant maximum over residues with on-beat occupancy tests whether phase invariance is what protects syncopation. Using raw Fourier magnitudes with no null calibration tests whether the null is needed at all.

Why not just…?

Why not just use a learned quality predictor like the SongEval-trained UTMOS or CMI-RM? This is the real threat. If these predictors already drop sharply on uniform reshuffles and collapsing maps, and stay flat on unit dilations, then our scorer adds interpretability and nothing else. The entanglement documented in SongEval, and CMI-RM's musicality–alignment SRCC near 0.85, suggest that learned scores mix factors. Even so, it is possible that they track meter. The perturbation audit settles this directly: we report each predictor's response per condition. If they match human responses, we reframe the contribution as the group-theoretic perception hypothesis and the test suite itself, not a new metric.

Why not just use beat-tracker confidence or inter-onset-interval statistics? Unit dilations change the sequence of inter-onset intervals while preserving class structure. A dilation by 5 mod 48, for example, reorders the gaps between hits. So these features and our score make opposite predictions on exactly the condition the hypothesis is about. Whichever agrees with listeners wins. If listeners reject dilated loops, interval statistics are vindicated and our invariance hypothesis is wrong. That outcome is still a clean finding about what meter perception requires.

Why not just estimate the null by Monte Carlo shuffling instead of generating functions? For moderate tails Monte Carlo works, and we expect the ablation to show near-identical rankings. The generating function earns its place in three ways. It gives exact, deterministic tails for strongly metrical loops, whose null probabilities are far below what feasible sampling resolves. It costs a polynomial power rather than millions of draws. And its roots-of-unity form yields the differentiable surrogate. We do not claim the scientific result depends on it.

Why not just run human listening panels? We do, but only to calibrate. As SongEval's 16 annotators and CMI-Pref's 31 experts show, expert panels do not scale to reranking or training loops, and their ratings are themselves entangled across dimensions.

Proposed Experiments

The decisive check is the invariance audit on FSLD and Looperman. The proposal predicts three things. Human ratings should satisfy identity ≈ translation ≈ unit dilation > within-coset shuffle > uniform reshuffle and collapse. FAD and IS should move little across all conditions. Our score should be invariant to unit dilation by construction and fall monotonically along the rest of that ordering. A significant human penalty for unit dilation relative to the identity re-render would falsify the hypothesis at its core.

The second experiment applies the scorer to StyleGAN2 loops. We sample loops from StyleGAN2 trained on Looperman and use the score to split them into high- and low-concentration halves. We then test whether humans rate the halves differently, even though both halves come from a model whose aggregate FAD looks competitive. This extends Hung et al.'s finding of no aggregate preference difference to a per-loop diagnostic.

The third experiment tests transfer to symbolic music, where tempo is not fixed. On MidiCaps outputs from PMT-4B, MIDI-LLM, and text2midi, which Agogic reports at Fréchet Music Distance (FMD) 196, 314, and 421 respectively, we compute the score after beat tracking. We then test whether PMT's 10 ms micro-timing lowers coarse-grid concentration. Here the phase-invariant, multi-level design either distinguishes expressive timing from disorder or fails visibly, and a small listening test decides which. FMD is reported throughout as the reference point.

Finally, the differentiable score is tested as a reranker. We use it for best-of-10 reranking of generated loops, mirroring the best-of-10 protocol that CMI-RewardBench reports as improving human A/B outcomes for CMI-RM. We compare it against CMI-RM reranking and against random selection under the same candidate pools.