"The question facing the music industry is no longer whether AI can generate convincing music. It can. The question is whether we can distinguish it — not by checking against a database of known AI outputs, but by reading the acoustic record of human biological limitation that every real musical performance leaves behind."
We define seven acoustic markers — collectively termed the Acoustic Humanity Index (AHI) — proposed to distinguish human-made music from AI-generated audio. Unlike existing detection approaches that pattern-match against known AI model outputs, the AHI framework is grounded in the acoustic consequences of human biological constraints: neuromuscular timing limits, respiratory cycles, inter-performer social adaptation, and cognitive chunking of musical time. We hypothesize that these constraints leave systematic acoustic fingerprints that AI generative systems, optimized for perceptual fidelity rather than biological authenticity, do not fully reproduce.
This is a method paper. It gives each marker a definition, a formula, a unit and a citation trail, and it specifies the evaluation protocol we will use. It does not report marker values, effect sizes, p-values or detection accuracy: no labeled human/AI evaluation set has been assembled yet, and the ZLMID v2.0 working database (20,405 tracks) currently holds metadata without acoustic enrichment. Each marker is grounded in established musicological and signal-processing literature, which makes the framework interpretable — but interpretability is not evidence of accuracy, and none is claimed here.
The index is defined in this paper. Evaluation on a labeled human/AI set is in progress. No accuracy figure exists yet, and none should be quoted from this work.
Deezer has publicly reported receiving on the order of 50,000 fully AI-generated tracks per day, roughly a third of its daily deliveries (Deezer newsroom, 2025). This is not a future scenario. It is the present operational reality for DSPs, rights organizations, and A&R departments.
The legal, economic, and cultural stakes are severe. Deezer has also reported that a large share of streams on fully AI-generated tracks are fraudulent — uploads intended to dilute royalty pools through synthetic stream farming. The Recording Industry Association of America has filed lawsuits against Suno and Udio alleging copyright infringement at scale, and major labels have since reached licensing settlements with AI music companies — the beginning of a regulated AI music ecosystem, but one that urgently requires reliable provenance verification.
Source: Deezer press releases on AI-generated uploads, 2025 — newsroom-deezer.com. Figures are Deezer's, not ours, and change month to month.
Most AI detection systems deployed today — from platform-internal classifiers to commercial detectors — operate on the same fundamental principle: train a model on known AI outputs, detect matches against that training distribution. This approach can perform well on the models it was trained on but contains a structural vulnerability: when a new generative architecture is deployed, detection degrades until the detector is retrained. Given how frequently AI music models are updated, this is an arms race that output-matching systems are architecturally disadvantaged in. Whether a biology-grounded approach does better is an empirical question this paper sets up but does not yet answer.
The Zeltor Labs Acoustic Humanity Index (AHI) proposes a different detection paradigm. Rather than asking "does this track match a known AI output?" it asks "does this track contain the acoustic signatures of biological human performance?"
Human music-making is biologically constrained in ways that are acoustically measurable. Humans cannot quantize perfectly. They breathe. They adapt to other performers. They slow down at phrase boundaries. They fatigue. They express. Our hypothesis is that these realities leave systematic acoustic fingerprints that are attenuated or absent in AI-generated audio — not because AI cannot simulate them individually, but because systems optimized for perceptual fidelity at the track level need not reproduce the full covariant structure of human performance.
If the hypothesis holds, the approach has a strategic advantage: it is architecture-agnostic. A new generative model may learn to simulate individual markers (near-perfect phrase-final lengthening, for example), but reproducing all seven simultaneously, with the correct covariant structure, would require encoding much of the biology of human musical performance. How durable that advantage is — and whether it exists at all — is exactly what the evaluation in Section IV is designed to test.
The composite AHI is a weighted sum of the seven per-marker scores, each normalized to 0–100. The current weights are hand-assigned priors and will be replaced by fitted weights once the labeled set exists.
Each marker below is a measurable acoustic quantity motivated by a specific aspect of human biological performance. Click any marker to expand its definition, formula, unit, working hypothesis and citation trail. Human and AI values for every marker are still to be measured; the qualitative directions stated are hypotheses drawn from the cited literature.
Human musicians exhibit systematic timing irregularities — deviations between intended and actual note onset times — arising from neuromuscular constraints, breath cycles, and inter-performer adaptation. The literature describes these deviations as long-range correlated stochastic processes (1/f-like noise). We hypothesize that they are attenuated in AI-generated audio, which tends toward near-perfect quantization.
Expected direction: higher in human recordings than in AI-generated tracks. Effect size, ranges and significance are to be measured on the labeled evaluation set described in Section IV.
This section will hold the measured results. Today it holds none: no labeled human/AI set has been assembled, no marker has been evaluated, and no accuracy has been computed. The protocol below is what will be run, so that readers can hold us to it.
A labeled evaluation set drawn from two sources: (a) human recordings with documented provenance — release metadata from the ZLMID v2.0 corpus, restricted to recordings whose performers and sessions are verifiable; (b) AI-generated tracks produced with named generators and recorded prompt/model/version. Genres will be stratified so no genre dominates either class. Target size and composition will be published with the results, not before.
Every track carries a binary label (human / AI) plus generator identity where applicable. The set is split into development and held-out test partitions before any marker is tuned. Marker normalizations and composite weights are fitted on the development partition only. The held-out partition is scored once. No figure from the held-out partition will be reported until that single pass is complete.
Per marker: distributions for each class, effect size, and separability (AUC). For the composite: accuracy, precision/recall per class, calibration, and performance broken out by genre and by generator — including generators absent from the development partition, which is the real test of architecture-agnosticism. Failures and null results will be reported alongside successes.
The ZLMID v2.0 working database holds 20,405 tracks of release and artist metadata drawn from MusicBrainz, AcousticBrainz and TheAudioDB. Acoustic enrichment (MFCC and spectral features) has not yet been run on any of them, so no marker can be computed at corpus scale today. The upstream sources are much larger than the working database; figures describing those sources are not counts of what has been analyzed here. Figures 1–5 of the earlier draft, which plotted marker values, accuracy comparisons, genre saturation and market projections, have been withdrawn because no measurement supported them.
If the markers separate the classes as hypothesized, the AHI could serve as one interpretable signal in an ingestion triage pipeline, alongside existing detectors. It should not be used as an automated rule, and no operating threshold is recommended until the held-out evaluation has been run and calibration is known.
Because each marker is a named, formula-defined acoustic quantity with a literature basis, the framework produces evidence that can be inspected marker by marker rather than a black-box score. That interpretability is the framework's contribution; it is not a claim about admissibility or legal weight, which this paper does not address.
Artists can run the seven heuristics on their own recordings today as a research preview and see the per-marker breakdown. The score is an experimental index, not a determination of authorship, and no certificate is issued. A verifiable statement of human authorship is the goal once the index has been measured.
The Acoustic Humanity Index described in this paper is implemented in Zeltor Labs AuthentiScan, a browser-based research preview. AuthentiScan runs seven acoustic heuristics on an uploaded file — audio never leaves the browser — and returns an experimental composite score with a marker-level breakdown. It issues no certificate, makes no determination of authorship, and is free to use while the index is being evaluated.
The music industry is facing a provenance problem unlike any it has encountered before. A large and growing share of new music arriving at distribution platforms may not have been made by a human being, and the economic, legal, and cultural consequences of that shift are accelerating.
The seven acoustic markers defined in this paper — the Acoustic Humanity Index — are a literature-grounded, architecture-agnostic proposal for responding to that problem. By measuring the acoustic consequences of human biological performance rather than the outputs of specific AI models, the framework aims at a detection approach that is durable and interpretable. Whether it is also accurate is not yet known.
To state it plainly: this paper defines the index. Evaluation on a labeled human/AI set is in progress, and no accuracy figure exists yet. The values, ratios, p-values and accuracy comparisons that appeared in the earlier draft of this paper were not backed by measurement and have been withdrawn. Results will be published here, under the protocol in Section IV, once the held-out partition has been scored.
The Zeltor Labs Institute for Music Intelligence will continue this work by first running acoustic enrichment on the ZLMID v2.0 working database and assembling the labeled set, then reporting the evaluation. Stem-level analysis (required for M6) and cross-market profiles are deferred until the base evaluation exists.