Table 1
The number of files is formed by a parallelized process with workers extracting form and chords data from a randomized folder of more than audio files, with a limit of JSON files.
| Collection | # Files | Duration Mean S.D. | Total Duration |
|---|---|---|---|
| Lastfm | 20,010 | 261.3 182.1s | 1452 h 6 m 56 s |
| Suno | 20,002 | 175.2 54.6s | 1084 h 47 m 9 s |
| Udio | 20,002 | 143.2 98.6s | 801 h 47 m 36 s |
Table 2
Summary statistics of likes and plays for the Suno () and Udio () collections, counting the tracks that completed the analysis pipeline. P90 and P95 denote the 90th and 95th percentiles.
| Metric | Collection | Mean | Med. | P90 | P95 | Max |
|---|---|---|---|---|---|---|
| Likes | Suno | 5.25 | 1 | 2 | 4 | 34,921 |
| Udio | 1.83 | 1 | 3 | 8 | 898 | |
| Plays | Suno | 204.88 | 4 | 26 | 50 | 445,490 |
| Udio | 44.10 | 4 | 32 | 58 | 48,940 |
Figure 1
Box plots of likes and plays for the Suno and Udio collections on a logarithmic scale. Songs from both platforms share a median of one like and four plays.
Table 3
Example output of consonance‑ACE. Each row contains the onset time, offset time, and estimated chord label. N denotes no chord detected. The model predicts a root pitch class and quality, labeling roots with sharps by default and assuming enharmonic equivalence.
| Start (s) | End (s) | Chord |
|---|---|---|
| 0.000 | 0.951 | N |
| 0.951 | 2.900 | G:min |
| 2.900 | 4.501 | F:maj |
| 4.501 | 7.471 | D#:maj |
| 7.471 | 9.118 | G:min |
| 9.118 | 11.346 | F:maj |
| 11.346 | 12.691 | C:min |
| 12.691 | 13.805 | D#:maj |
| 13.805 | 15.360 | G:min |
| 15.360 | 16.914 | F:maj |
| 16.914 | 18.840 | D#:maj6 |
| 18.840 | 21.787 | G:min |
| 21.787 | 23.179 | F:maj |
| 23.179 | 24.617 | D#:maj |
| 24.617 | 25.499 | G:min |
| 25.499 | 26.241 | D#:maj |
Figure 2
Top 30 most common trigrams found in the Lastfm, Suno, and Udio datasets. The y‑axis shows the relative frequency of each trigram in the collection. The distribution of styles differs across collections.
Figure 3
Vocabulary distribution of chord qualities over the collections. The y‑axis indicates the total number of occurrences for each chord quality in each collection, in units of occurrences.
Figure 4
Distribution of most common tonalities in all three collections.
Figure 5
Per‑style harmonic palette. For each style, the axes are the most frequent chord trigrams (ordered by Lastfm), and each collection is scaled to its own maximum within that style; the figure therefore shows which progressions each collection draws on, not how often. Polygon size carries no magnitude. Absolute frequencies appear in Figure 6 and Table 4.
Figure 6
Relative frequency of canonical pop/rock trigrams in Suno compared with Lastfm and Udio. Bars show the relative frequency of each progression within a collection, and all differences between Suno and the reference collections are significant at .
Table 4
Suno‑enriched trigrams. Panel A lists the most prevalent canonical progressions, defined as any permutation of chords drawn exclusively from {I/i, IV/iv, V/v, vi/VI/bVI}. Panel B lists the highest RR non‑canonical trigrams. = raw occurrence count; RR = risk ratio (Suno/Lastfm). All rows are statistically significant (FDR‑adjusted ).
| Trigram | % Suno | % Udio | % Last | RR | |||
|---|---|---|---|---|---|---|---|
| Panel A: Top 15 canonical enriched trigrams | |||||||
| V‑vi‑IV | 40007 | 10782 | 11629 | 2.896 | 1.165 | 0.834 | 3.47 |
| vi‑IV‑I | 39031 | 8725 | 11163 | 2.825 | 0.943 | 0.801 | 3.53 |
| IV‑I‑V | 32875 | 10697 | 14674 | 2.379 | 1.156 | 1.053 | 2.26 |
| vi‑IV‑V | 25877 | 9385 | 10198 | 1.873 | 1.014 | 0.732 | 2.56 |
| I‑V‑vi | 25364 | 4823 | 6003 | 1.836 | 0.521 | 0.431 | 4.26 |
| IV‑V‑vi | 25068 | 10078 | 10618 | 1.814 | 1.089 | 0.762 | 2.38 |
| I‑vi‑IV | 12318 | 3988 | 5471 | 0.892 | 0.431 | 0.393 | 2.27 |
| IV‑I‑vi | 11761 | 4051 | 5851 | 0.851 | 0.438 | 0.420 | 2.03 |
| vi‑V‑IV | 10190 | 2837 | 4146 | 0.738 | 0.307 | 0.297 | 2.48 |
| V‑I‑vi | 9768 | 3394 | 4252 | 0.707 | 0.367 | 0.305 | 2.32 |
| V‑i‑bVI | 8980 | 2609 | 2165 | 0.650 | 0.282 | 0.155 | 4.18 |
| vi‑V‑I | 7073 | 2371 | 3261 | 0.512 | 0.256 | 0.234 | 2.19 |
| v‑i‑bVI | 6910 | 1626 | 1762 | 0.500 | 0.176 | 0.126 | 3.96 |
| IV‑vi‑V | 6810 | 2129 | 3143 | 0.493 | 0.230 | 0.225 | 2.19 |
| bVI‑V‑i | 6741 | 2452 | 2211 | 0.488 | 0.265 | 0.159 | 3.08 |
| Panel B: Top 10 non‑canonical enriched trigrams (by RR) | |||||||
| iv‑ii‑v | 96 | 19 | 10 | 0.007 | 0.002 | 0.001 | 9.69 |
| iii‑vi‑ii | 1694 | 280 | 202 | 0.123 | 0.030 | 0.014 | 8.46 |
| I‑bvi‑bVII | 92 | 23 | 13 | 0.007 | 0.002 | 0.001 | 7.14 |
| II‑v‑III | 245 | 32 | 35 | 0.018 | 0.003 | 0.003 | 7.06 |
| bVI‑#iv‑V | 398 | 78 | 75 | 0.029 | 0.008 | 0.005 | 5.35 |
| iii‑ii‑v | 79 | 22 | 16 | 0.006 | 0.002 | 0.001 | 4.98 |
| III‑bvii‑i | 270 | 81 | 55 | 0.020 | 0.009 | 0.004 | 4.95 |
| IV‑iii‑vi | 1859 | 462 | 384 | 0.135 | 0.050 | 0.028 | 4.88 |
| V‑I‑bvi | 111 | 26 | 23 | 0.008 | 0.003 | 0.002 | 4.87 |
| vi‑ii‑v | 771 | 200 | 165 | 0.056 | 0.022 | 0.012 | 4.71 |
Figure 7
Section‑label tetragram heatmap across collections. Each row is a tetragram of four consecutive section labels, extracted after collapsing repeated adjacent sections. Cell shading encodes relative frequency within each collection; integer annotations report absolute counts. Tetragrams are sorted by descending Lastfm frequency.
Table 5
Proportion of songs (%) per collection whose Axis rate exceeds each threshold, under the two measures: the four chords in any order (72 patterns) and the loop only (12 rotations). Collections: Lastfm (), Suno (), Udio (), counting songs with at least one four‑chord window.
| Axis rate | Any order (72) | Loop only (12) | ||||
|---|---|---|---|---|---|---|
| L | S | U | L | S | U | |
| 0% | 26.9 | 47.2 | 27.1 | 10.6 | 27.5 | 10.6 |
| 1% | 24.9 | 46.2 | 26.3 | 9.3 | 26.4 | 9.8 |
| 5% | 13.5 | 35.2 | 16.1 | 4.1 | 17.5 | 5.0 |
| 10% | 8.6 | 28.0 | 10.5 | 2.8 | 13.6 | 3.4 |
| 25% | 3.5 | 16.1 | 4.4 | 1.3 | 8.3 | 1.7 |
| 50% | 0.8 | 6.5 | 1.3 | 0.3 | 3.8 | 0.6 |
Figure 8
Normalized musical form trigram profiles across eight human‑annotated genre tags per collection. Each axis represents a trigram of three consecutive section labels (e.g., chorus–verse–chorus), extracted after collapsing repeated adjacent sections. For each genre, the top trigrams are selected by Lastfm frequency, and values are normalized to the maximum count per collection within each genre.
Figure 9
Most frequent harmonic tetragrams per collection. Darker cells indicate stronger prevalence, and integer annotations report absolute counts. The left panel is ranked by Suno, whose top patterns are the rotations of the Axis loop (e.g. vi‑IV‑I‑V, I‑V‑vi‑IV); the right panel is ranked by Udio, whose top patterns are the repetition progressions I‑IV‑I‑IV, IV‑I‑IV‑I, and i‑bVI‑i‑bVI, all very common in human‑composed music, where Lastfm’s counts are highest.
Figure 10
Per‑song Axis rate under the two measures, for songs with at least 10% Axis content. (a) Any order: a four‑chord window counts when its chords form an Axis set in any order (72 patterns across the three sets). (b) The loop: a window counts only when it is one of the 12 rotations of the three loops (four per set). Curves are kernel‑density estimates; each legend percentage is the share of that collection’s songs above the 10% threshold under that measure.
Figure 11
Loop versus reordering. Among four‑chord windows built on the Axis chords, the loop (the chords in cyclic order) is separated from the same chords in a different order. (a) Each as a share of all four‑chord windows, loop (solid) and reordering (hatched); the number inside the solid segment is the loop share alone, and the number above each bar is the combined total. (b) The loop’s share of those Axis‑chord windows. The dashed line marks 50%.
Figure 12
Distribution of song durations across the three collections (Lastfm, Suno, Udio). The x‑axis represents duration in seconds, and the y‑axis indicates the number of songs per 10‑s bin.