Data‑Driven Analysis of Musical Form and Harmonic Structure in AI‑Generated Popular Music: A Case Study with Suno and Udio

2.1 AI music generation and evaluation

Recent open‑source work on text‑conditioned music and audio synthesis includes models such as MusicGen (Copet et al., 2023), MusicLM (Agostinelli et al., 2023), and Stable Audio Open (Evans et al., 2025), representing the state of the art in publicly available generative audio research. However, the focus of this paper is not on the generative models themselves but on the musical properties of their output, analyzed at the corpus scale in comparison with human‑produced music.

Sturm and Ben‑Tal (2017) evaluates folk‑rnn, a generative system trained on a large corpus of folk tunes in ABC notation, examining outputs in relation to musical practice rather than only automatic metrics, with a concrete evaluation pipeline, corpus statistics analysis, and focused musical inspection to locate where model outputs resemble and where they diverge from human practice. Ariza (2009) argues that discrimination listening tests, lacking the interrogation central to the Turing test they are often framed as, at most establish that a digital music artifact sounds plausible to listeners, which is insufficient to characterize the musical quality produced by an AI system.

Dervakos et al. (2021) propose a framework for evaluating symbolic AI‑generated music using tools from music theory, defining heuristic metrics grounded in tonic coordinate systems and tone networks, such as the circle of fifths, to quantify musicality without relying on user surveys or real data. Wu and Yang (2020) analyze the outputs of the Jazz Transformer using objective metrics, including pitch‑class histogram entropy, chord‑progression irregularity, groove‑pattern similarity, and structureness indicators, demonstrating that systematic shortcomings in AI‑composed music are not fully captured by listening tests alone. Notably, their definition of chord progression irregularity as the proportion of unique chord trigrams in a piece establishes a precedent for trigram‑based harmonic analysis of AI‑generated music.

Some studies find biased perception of music when works are labeled as AI (Déguernel and Sturm, 2023; Ragot et al., 2020; Shank et al., 2022), while others report null effects (Pasquier et al., 2016; Zlatkov et al., 2023). Together, these motivate a careful, data‑driven comparison between music made with current AI platforms and music created by humans.

2.2 Music structure analysis

MSA is the task of segmenting an audio recording into contiguous, non‑overlapping sections and assigning labels that reflect musical function or similarity (Nieto et al., 2020). Computational approaches exploit four main principles: homogeneity (sections internally consistent with respect to some musical attribute), repetition (detecting recurrent sequences), novelty (identifying boundary points where musical content changes abruptly), and regularity (statistical properties of segment durations). Classical methods relied on self‑similarity matrices and novelty curves derived from hand‑crafted features such as chromagrams and Mel‑frequency cepstral coefficients (MFCCs) (Foote, 2000), repetition and recurrence‑based representations (Levy and Sandler, 2008; Serrà et al., 2012b), Laplacian spectral decomposition (McFee and Ellis, 2014), and barwise self‑similarity with dynamic programming (Marmoret et al., 2023). Despite these advances, MSA remains an open challenge due to the subjectivity of structural annotations: the best reported structure labeling scores reach approximately $74\%$ on the Harmonix Set (Kim and Nam, 2023), well below the estimated human inter‑rater agreement of approximately $90\%$ (Nieto et al., 2020).

More recent approaches have improved structural analysis through deep learning, including few‑shot embeddings for Laplacian decomposition (Salamon et al., 2021), unsupervised contrastive multi‑level representations (Buisson et al., 2022), self‑supervised novelty detection (Peeters, 2023), hierarchical spectral–temporal Transformers for direct semantic labeling (Wang et al., 2022), and graph attention networks for pairwise structure prediction (Buisson et al., 2024).

For functional structure extraction, we employ the All‑In‑One model (Kim and Nam, 2023), which jointly performs beat tracking, downbeat detection, segment boundary detection, and semantic section labeling (verse, chorus, bridge, etc.) in a single architecture. We adopt it for two reasons: it directly outputs the section labels our analysis requires, and it reaches state‑of‑the‑art performance on the Harmonix Set across all four tasks.

2.3 Automatic chord estimation

ACE is a central task in MIR with applications in music performance, transcription, and formal analysis research. ACE has evolved from early template and chroma‑based systems with probabilistic sequence models, such as hidden Markov models, conditional random fields, or deep belief networks (Humphrey and Bello, 2015; Mauch and Dixon, 2009; Ni et al., 2012), to more recent deep learning approaches that learn end‑to‑end representations from spectral features, increasingly targeting large chord vocabularies that include sevenths, extensions, and inversions.

Large‑vocabulary ACE has advanced through several complementary strategies. McFee and Bello (2017) decouple pitch‑class detection from label assignment in a 170‑class framework; Jiang et al. (2019) decompose chord labels into independent musical components (root, quality, bass, extensions), generalizing to 301+ classes; and Park et al. (2019) propose BTC, a bi‑directional Transformer achieving competitive large‑vocabulary performance with publicly available code. Most recently, Poltronieri et al. (2025) propose consonance‑ACE, a Conformer‑based model whose decomposed decoder reconstructs chord labels without a fixed vocabulary, paired with consonance‑based label smoothing, which outperformed BTC on both standard and non‑binary evaluation metrics on RWC Pop and USPop. Waseem Akram et al. (2025) further extends this line with ChordFormer for large‑vocabulary recognition and usability.

Despite these advances, implementations remain limited in coverage. Open‑source libraries such as Madmom (Böck et al., 2016) and librosa (McFee et al., 2025) implement baseline chord‑estimation algorithms, most of which remain fundamentally limited to detecting major and minor triads, struggling with tetrads, extensions, suspensions, and inversions. Earlier efforts to expand chord vocabularies, such as Chordino (Mauch and Dixon, 2010), are no longer maintained. Furthermore, analyses relying on chord‑class aggregation can misrepresent stylistic progressions: Collins (2025) demonstrates that Chordino‑based histograms failed to capture modal‑interchange progressions while detecting only broad palette differences. Multi‑source approaches such as DECIBEL (Odekerken et al., 2021), which fuse audio features with crowd‑sourced MIDI annotations and guitar tablature, partially address vocabulary restrictions, yielding improvements over audio‑only baselines, though major and minor accuracy typically remain between $75\%$ and $87\%$, a ceiling attributed to annotation subjectivity.

Roman numeral analysis (RNA) encodes chords relative to a tonal center. Benchmarks on the PARC dataset (Poppe et al., 2025) confirm that models predicting Roman numerals directly from audio achieve modest accuracy, particularly for complex progressions. We therefore derive Roman numerals in two stages rather than directly from audio: ACE recognizes the chords, and a symbolic step assigns each its key‑relative function through global key estimation (Section 4.1).

Given these considerations, this study adopts consonance‑ACE (Poltronieri et al., 2025) for three reasons: (1) its decomposed decoder is not constrained to a fixed vocabulary, accommodating harmonic diversity beyond major/minor triads; (2) consonance‑based label smoothing produces errors that fall between harmonically related chords rather than random ones; and (3) its demonstrated superiority over BTC on standard benchmarks, combined with publicly available code, provides a reproducible foundation for large‑scale analysis. The model is described in detail in Section 4.1.

2.4 Harmonic analysis of corpora

Several annotated corpora and computational analyses have been dedicated to the study of harmony and musical structure in popular music. The hand‑annotated rock corpus of Temperley and Clercq (2013) supports statistical investigation of chords, scale degrees, and key‑finding models. Expert ground‑truth chord annotations for charting songs support research in automatic chord recognition and music analysis (Burgoyne et al., 2011). Computational analysis of Billboard chart data has been used to study the evolution of harmony, timbre, and style in popular music from 1960 to 2010 (Mauch et al., 2015). Long‑term trends in harmonic and timbral features of contemporary Western popular music have also been examined to document changes in musical diversity (Serrà et al., 2012a). Automated approaches to chord and key estimation provide frameworks for the simultaneous identification of chords and musical context from audio data (Mauch and Dixon, 2009). There is no comprehensive corpus‑level comparison of harmonic and structural properties between music generated by current AI platforms and human music.

4.1 ACE

Chord estimation is performed using consonance‑ACE (Poltronieri et al., 2025), which independently predicts root, bass, and pitch‑class activations and reconstructs chord labels symbolically through a decomposed decoder. The model operates on CQT spectrograms computed over 20‑s chunks and outputs frame‑level chord predictions with no constraint to a fixed vocabulary. We use the publicly available pretrained model without modification.¹¹ Chord segments shorter than $0.5$ s are merged into adjacent segments, and consecutive frames with identical chord labels are collapsed into continuous intervals.

The output of the chord estimator consists of absolute chord labels (e.g., D:min7, G:7, C:maj7) aligned with time onset‑offset columns, as shown in Table 3. To enable key‑relative analysis, we convert these to functional harmony representations (e.g., ii, V, I). For each track, we first estimate a single global key from the chord sequence by duration‑weighted diatonic template matching: each of the $12$ major and $12$ minor keys is scored by the time spent on chords whose root and quality fit its scale degrees. The comparison operates on pitch classes, so the sharp‑based spelling of the chord labels (Table 3) plays no role: D# and Eb count as the same root, and the estimated key is spelled conventionally (E$♭$ major rather than D$♯$ major). Each absolute chord label is then mapped to its corresponding Roman numeral degree relative to the estimated key using the music21 library (Cuthbert and Ariza, 2010). Consecutive identical Roman numerals are de‑duplicated before downstream n‑gram analysis. The resulting chord intervals are subsequently aligned to beat positions estimated by the All‑In‑One model (Section 4.2), yielding the beat‑level chord annotations.

4.2 Musical form estimation

Musical structure is extracted using the All‑In‑One model (Kim and Nam, 2023), which jointly performs four tasks in a single architecture: beat tracking (locating regular pulse positions), downbeat detection (identifying the first beat of each bar), segment boundary detection (finding time points where sections change), and functional structure labeling (assigning semantic labels such as verse, chorus, bridge, intro, outro, or solo to each segment).

The model first separates the input audio into four stems (drums, bass, vocals, other) using Hybrid Transformer Demucs (Rouard et al., 2023) and processes the resulting spectrograms through a stack of transformer modules combining dilated and neighborhood attention. Four output heads produce frame‑level predictions for each task. The model achieves state‑of‑the‑art performance on the Harmonix Set ($F1=0.958$ for beats, $F1=0.915$ for downbeats, $HR.5F=0.660$ for segmentation, $PWF=0.738$ for structure labeling).

We apply the publicly available pretrained model without modification. For each track, the model outputs a sequence of labeled sections with their onset and offset times, stored as _musical_form.lab files. Consecutive identical section labels are subsequently merged before downstream analysis.

4.3 N‑gram dataset formation

To systematically analyze harmonic tendencies across collections, we construct two n‑gram datasets: a trigram dataset of all consecutive three‑chord sequences and a tetragram dataset of all consecutive four‑chord sequences from each collection. For example, the progression $Dm7$, $Ab7$, $Gm7$, $C{\#}^{\circ }7$, $Dm7$ from a Pop song (‘Azul’ by Djavan), yields the trigrams $[Dm7,Ab7,Gm7]$, $[Ab7,Gm7,C{\#}^{\circ }7]$, and $[Gm7,C{\#}^{\circ }7,Dm7]$; the example uses absolute chord labels for illustrative clarity.

Both datasets are formed from the same normalized sequences. Absolute chord labels are converted to Roman numerals as described in Section 4.1, and ‘no‑chord’ labels are excluded. Each Roman numeral is then respelled onto a single 12‑tone vocabulary, with one name per pitch class: #V becomes $♭$VI, and only the tritone degree keeps a sharp spelling (#IV or #iv). Consecutive identical Roman numerals are de‑duplicated. For each song, a sliding window of $n$ chords

is passed over the Roman numeral sequence, where $N$ is the number of chord events after de‑duplication, $C_i$ denotes the $i^{\text{th}}$ Roman numeral, and $n=3$ for trigrams or $n=4$ for tetragrams.

All windows enter both datasets, including repetition‑based patterns such as I‑IV‑I‑IV and sequences spanning section boundaries; windows built on a single root are excluded, and songs with fewer than $n$ chords after de‑duplication yield no windows. N‑grams are counted per collection, revealing the most common progressions and their cross‑collection frequency distributions.

4.4 ‘Axis of Awesome’ chord progression analysis

We measure how often songs in each collection contain chord progressions matching the Axis of Awesome (Richards, 2017), using the Roman numeral sequences produced by the ACE pipeline.

The Axis of Awesome is a specific progression, not merely a set of chords: the four chords {I, V, vi, IV}, played as a repeating cycle in that order, with a minor form {i, $♭$VI, III, $♭$VII} and a minor variant {i, $♭$VI, $♭$III, $♭$VII}. Pop harmony also draws on these chords in other orders, and our first inspections of the Suno collection showed a marked concentration on them; we therefore measure this tendency systematically, with two measures that separate using the chords from playing the progression. The first, Axis‑family membership, counts a four‑chord window when its chords form one of the three sets in any order ($24$ possible orderings of four chords, $72$ patterns across the three sets) and asks whether a track draws on the Axis chord material at all. The second, the literal Axis loop, counts a window only when its chords follow the cycle’s order. A cycle has no fixed starting point, so I‑V‑vi‑IV, V‑vi‑IV‑I, vi‑IV‑I‑V, and IV‑I‑V‑vi are the same loop entered at different chords; each set thus contributes $4$ rotations, $12$ patterns in total, and this measure asks whether the track plays the progression itself. The match is per window: a window counts when its four chords follow the loop’s cyclic order, whether or not the song repeats it; sustained cycling shows up as a high per‑song rate. Richards (2017) observes that the cycle strongly tends to appear in its given order. Rotation equivalence also absorbs relative‑major/minor ambiguity in the global key estimate from which the Roman numerals are derived (Section 4.1): the four chords C, G, Am, and F are labeled I‑V‑vi‑IV when the key is read as C major but III‑bVII‑i‑bVI, the minor‑form set begun on a different chord, when it is read as the relative A minor. Treating all three sets and their rotations as one loop recovers the progression in both readings, so a track is not counted as non‑Axis merely because its key estimate landed on the relative key. From here, we will refer to this pattern of chord progressions as Axis.

A sliding window of four chords is passed over each deduplicated Roman numeral sequence, and every four‑chord window $(a,b,c,d)$ is tested for membership in both pattern sets: the $72$ Axis‑family patterns and the $12$ rotation‑only patterns.

The tetragram Axis rate for a song is the proportion of its four‑chord windows that match, expressed as a percentage; we compute it separately under the Axis‑family and the rotation‑only measures.

5.1 Trigram tendencies across collections

Figure 2 shows the 30 most prevalent trigram patterns in the three collections (Lastfm, Suno, and Udio), distributed over the 10 most common styles found in the metadata descriptors of the songs. The label ‘Others’ collects several styles in one group. Both Udio and Suno show a strong tendency toward Pop, Rock, Metal, and Electronic.

Figure 2

While we acknowledge that the style tags associated with tracks in our dataset are inherently flawed and subject to user bias (users may arbitrarily assign genres, and the platforms’ own tagging systems lack rigorous musicological grounding), we use these classifications, as they reflect the styles users most frequently specify when generating music. As demonstrated by Casini et al. (2026), these same style descriptors (Pop, Rock, Electronic, Hip‑hop, etc.) are among the most prevalent tags in the dataset.

The collection‑level differences visible in Figure 2 are quantified in detail in Sections 5.2 and 5.4.

Figure 3 shows the distribution of chord qualities, which is similar across the three collections. The major chord is the most common, followed by minor, power chords, suspended fourths, and dominant sevenths. Altered dominants, dominant chords with a sharp 9th or flat 13th, appear in both modes and function as modal interchange in major keys. The dominant sharp 11th, once ubiquitous in jazz, bossa nova, and soul, is rare in all three collections. Power chords, which lack the third that defines mode, appear in both major and minor keys.

Figure 3

Figure 4 shows the distribution of tonalities in each collection, ordered by their frequency in Lastfm. The proportions are broadly similar across the three collections; Suno, however, favors C major, D major, G major, and B major more strongly than Lastfm and Udio do, while Udio shows a stronger presence of A minor, G minor, and Eb minor. The distribution of tonalities is consistent with the circle of fifths.

Figure 4

Figure 5 shows the harmonic palette of each collection across the eight most common styles (Rock, Electronic, Pop, Jazz, Country, Hip‑Hop, Soul, and Blues): for every style, it shows the most frequent chord trigrams and how each collection distributes its usage across them. To compare the palettes by shape, each collection is normalized to its own maximum within a style, so the figure describes which progressions a collection draws on, not how often it uses them. Read this way, the three collections share a common canonical core but differ in emphasis.

Figure 5

5.2 Harmonic tendencies

Figure 6 shows the relative frequency of canonical pop/rock chord trigrams, defined as permutations of {I/i, IV/iv, V/v, vi/VI/$♭$VI}, in each collection. The figure demonstrates that Suno tends to favor chord progressions strongly associated with the Western pop idiom, particularly those prevalent since the early 2000s. For every progression tested, Suno’s relative frequency is significantly greater than that observed in the other collections ($p<0.001$ for all comparisons). As shown in Temperley and Clercq (2013), the centrality of progressions such as I‑V‑vi‑IV, IV‑I‑V, and related idioms is a defining feature of Western pop/rock music; they report that the most frequent non‑tonic chord roots are IV, V, bVII, and vi, with progressions among these chords forming the core stylistic harmonic vocabulary. Our results both replicate and extend these findings, revealing the extent to which such progressions are even more prevalent in AI‑generated music from Suno.

Figure 6

One normalization precedes this analysis: the four Aeolian‑cadence spellings of a progression are folded into their relative‑major equivalents (e.g. i‑bVI‑III is counted with vi‑IV‑I), so that the same progression is not split between its major and relative‑minor readings.

To quantify the tendencies shown in Figure 6 for each canonical progression in Suno relative to the reference collections (Lastfm and Udio), we calculated a risk ratio (RR) for every trigram. Raw counts are misleading because Suno contains far more trigrams than either reference collection, so the RR normalizes by each collection’s total:

where $N_{\mathrm{Suno},t}$ and $N_{\mathrm{Ref},t}$ are the counts of trigram $t$ in Suno and in the reference collection, respectively, while $N_{\mathrm{Suno}}$ and $N_{\mathrm{Ref}}$ denote the total numbers of trigrams in those two collections.

To test whether the difference is greater than expected by chance, we built a $2\times 2$ contingency table (Suno vs. Reference $\times$ $t$ vs. ‘all other trigrams’) and applied Fisher’s exact test. All raw $p$‑values are corrected for multiple comparisons with the Benjamini–Hochberg false‑discovery‑rate procedure (Benjamini and Hochberg, 1995); the resulting $p$‑value is used to filter and choose the trigrams that are shown in Table 4:

• at least $30$ occurrences in Suno and $10$ in each reference collection,

• risk ratio $RR\ge 2$,

• adjusted $p-$value $<0.05$.

where $a=N_{\text{Suno},t}$, $b=N_{\text{Suno}}-a$, $c=N_{\text{Ref},t}$, and $d=N_{\text{Ref}}-c$. A risk ratio $R{R}_t>1$ together with $p<0.05$ indicates that trigram $t$ is significantly more common in Suno than in the reference collection, as is the case for all the trigrams shown.

Table 4 shows a representative selection of the $252$ Suno‑enriched trigrams passing all filters: Panel A lists the top 15 by Suno count of the 95 canonical progressions, while Panel B lists the top 10 non‑canonical trigrams by RR from the remaining 157. The columns provide the raw counts in Suno ($C_{\text{Suno}}$), Udio ($C_{\text{Udio}}$), and Lastfm ($C_{\text{Last}}$); their percentage of the total vocabulary; and the RR relative to Lastfm.

Across all canonical progressions in Panel A, RRs range from $2.0\times$ to $4.3\times$, with the major‑mode Axis chains (e.g., I‑V‑vi at $4.3\times$, vi‑IV‑I at $3.5\times$) among the most enriched. Every canonical pattern is significantly enriched ($\mathtt{p-adj}<10^{-16}$).

5.3 N‑Gram analysis of section‑label sequences

We analyze sequences of section labels to explore patterns in song structure, using the same n‑gram construction applied to chords (Section 4.3). For each track, the section sequence is read from the All‑In‑One output, consecutive identical sections are collapsed, and trigrams and tetragrams of consecutive section labels are extracted with a sliding window (e.g. verse–chorus–verse).

This yields $104,865$ section‑label trigrams in Lastfm, $71,870$ in Suno, and $56,232$ in Udio ($232,967$ in total), together with the corresponding tetragrams ($86,086$, $53,704$, and $42,751$, respectively). Their cross‑collection frequencies reveal the structural motifs that reflect how songs are typically organized.

Figure 7 shows the most frequent section‑label tetragrams; the results are discussed in Section 6.3.

Figure 7

As a collection‑level view of formal diversity, we count the distinct full song forms in each collection: each track’s section sequence, after collapsing consecutive repeats, is one form. The three size‑matched collections collapse to $3,349$ unique forms in Suno, $3,814$ in Udio, and $6,020$ in Lastfm.

5.4 Axis of Awesome tendencies

We next measure how strongly each collection leans on the Axis of Awesome. Table 5 reports the proportion of songs whose Axis rate, the share of their four‑chord windows that match (Section 4.4), exceeds a range of thresholds.

At the $\ge$10% threshold, counting the four chords in any order, $8.6\%$ of Lastfm, $10.5\%$ of Udio, and $28.0\%$ of Suno songs qualify as Axis‑based; counting only the loop, the rates are $2.8\%$, $3.4\%$, and $13.6\%$ (Table 5). Figure 8 shows the distribution of per‑song rates behind these counts. Given the collection sizes, the $95\%$ confidence interval of any proportion in Table 5 is below $\pm 0.8$ percentage points, so the differences between Suno and the other collections far exceed sampling uncertainty. Both measures confirm the findings of Richards (2017) on the cultural ubiquity of this progression and extend them to show that Suno reproduces it at a substantially higher rate than both the human reference collection and Udio, with the lead widening when only the loop counts (a Suno‑to‑Lastfm ratio of $3.2\times$ rising to $4.8\times$).

Figure 8

The rates above count how many songs lean on the Axis; they do not say, when a collection reaches for these four chords, whether it plays the loop or rearranges them. Figure 9 answers that, splitting the four‑chord windows built on the Axis chords into the loop itself and the same chords in a different order. Panel (a) gives each as a share of all four‑chord windows: Suno places 12.4% on the Axis chords (3.5% Lastfm, 4.6% Udio), most of it in the loop. Panel (b) is the loop’s share of that usage: 54.3% for Suno, against 36.3% for Lastfm and 38.5% for Udio. When Suno uses these chords, it usually plays the loop rather than rearranging it.

Figure 9

6.1 Applying MIR tools to large music corpora

The chord estimator and structural model adopted in this study address the principal methodological concerns for corpus‑scale analysis, but challenges remain that should be acknowledged. For chord estimation, even state‑of‑the‑art models exhibit lower accuracy on complex chord types (sevenths, extensions, suspensions, inversions) than on triads (Pauwels and Peeters, 2013; Poltronieri et al., 2025), and reported performance is sensitive to evaluation vocabulary and chord‑mapping strategies. The conversion from chord labels to Roman numerals depends on a single global key estimated per track, which can be wrong under modulation, strong modal mixture, or relative‑major/minor ambiguity and then propagates to every Roman numeral in the track. The same estimator runs identically on all three collections, so any such error is shared and cannot create the between‑collection differences we report. Furthermore, consonance‑ACE assumes enharmonic equivalence: sharp 9ths, sharp 11ths, and flat 13ths are encoded as their enharmonic equivalents (flat third, flat fifth, and flat sixth, respectively), and therefore do not appear as distinct labels in the chord vocabulary.

This study operates at the level of chord‑label distributions and n‑gram sequential patterns, which is the appropriate granularity for corpus‑scale comparison across $60,014$ tracks. Voice leading, harmonic rhythm, and chord‑voicing detail are not captured in any chord‑symbol–based corpus study and fall outside the present scope.

For structural analysis, the All‑In‑One model provides semantic section labels (verse, chorus, bridge, intro, outro) unavailable from unsupervised methods. Its reported scores on the Harmonix Set ($HR.5F=0.660$ for segmentation, $PWF=0.738$ for structure labeling) remain below estimated human inter‑rater agreement (Nieto et al., 2020), so a degree of error is present in the annotations. However, the same model is applied uniformly across all three collections, so tool choice cannot account for between‑collection differences, and a labeling artifact triggered by AI‑generated audio as such would push Suno and Udio in the same direction, whereas they deviate from Lastfm in opposite directions on specific patterns: Suno closes a verse‑chorus cycle with an outro at more than twice the human per‑song rate versus Udio does so at two‑thirds of it (Section 6.3). Our analysis targets distributional tendencies across $60,014$ tracks; systematic differences between collections therefore reflect genuine structural tendencies rather than tool‑specific artifacts.

Platform genre tags record the surface aesthetic a user asks for (timbre, mood, instrumentation), not harmonic function, and the two seem decoupled: the ‘Jazz’ tag on Suno and Udio returns mostly ambient instrumental tracks, without the ii‑V‑I cadential vocabulary that defines part of the jazz repertoire. The per‑genre panels (Figure 5) therefore report what each platform files under a label, not a fixed musicological category, and our corpus‑level contrasts do not depend on this: they are computed across all tracks regardless of tag.

Despite these caveats, the combination of a chord estimator with a semantically aware structure analyzer represents a meaningful advance over the triad‑level, label‑free pipelines that were the practical default for large‑scale audio analysis until recently.

6.2 Harmonic comparison of AI‑generated and human‑composed music

The four metrics used in this section capture different dimensions of harmonic practice and should be read accordingly. Chord‑quality distributions (Figure 3) describe the harmonic palette, which chord types appear and how often. Trigram frequencies and cross‑collection correlations describe local sequential behavior, in which three‑chord successions recur and at what relative rates. Axis rates measure use of a canonical four‑chord set, reported both as family membership (any ordering) and as the literal loop (the canonical rotation order). Tetragram frequencies describe recurring four‑chord patterns and how often they recur within a collection. These dimensions are complementary but independent: a high trigram‑frequency correlation between two collections, such as the $r=0.981$ shared by Udio and Lastfm, indicates distributional similarity in local harmonic motion but says nothing about whether those progressions occupy the same formal positions within a song or occur at comparable rates of harmonic change.

Three strands of evidence converge on a consistent picture of how AI‑generated harmony relates to human‑composed music.

At the level of chord‑quality vocabulary, the three collections are broadly similar: major triads dominate, followed by minor triads, power chords, seventh chords, dominant sevenths, and suspended chords (Figure 3). This indicates that current generative platforms reproduce the surface‑level harmonic palette of popular music with fidelity. The differences emerge not in what chord types appear, but in how they are sequenced.

Trigram frequency analysis reveals that Suno concentrates its harmonic motion on a narrow subset of Western pop progressions. Across all 95 canonical progressions that pass the enrichment filters (FDR‑adjusted $p<0.05$, RR $\ge 2.0$), Suno is overrepresented at frequency ratios ranging from $2.0\times$ to $7.9\times$ relative to Lastfm. The highest‑frequency 15 of these 95 progressions, listed in Panel A of Table 4, concentrate the over‑representation at $2.0\times$–$4.3\times$, while the remaining lower‑frequency canonical progressions contribute ratios between $2.0\times$ and $7.9\times$. The cumulative effect across all 95 enriched progressions is substantial: Suno’s harmonic vocabulary is contracted to $9,157$ unique trigram types, compared to $11,115$ in Udio and $11,546$ in Lastfm.

At the four‑chord level, the Axis of Awesome analysis (Section 5.4) sharpens the same contrast. Suno is Axis‑based far more often than either human‑composed music or Udio, and the gap widens when matching is restricted to the literal loop rather than any ordering of the four chords: the order‑sensitive test does not soften Suno’s lead but increases it at every threshold. Suno does not merely share the Axis vocabulary; it plays the loop in its canonical order far more often than human music does.

Udio’s harmonic distributions, by contrast, are statistically closer to those of Lastfm across all three measures. Cross‑collection trigram frequency correlations confirm this: Udio and Lastfm share $r=0.981$, while Suno correlates with Lastfm at only $r=0.795$. The harmonic tetragram frequencies (Figure 10) reinforce this at the four‑chord level: Udio and Lastfm share a similar tetragram distribution ($r=0.967$) while Suno diverges toward a smaller set of high‑frequency patterns ($r=0.662$ with Lastfm). As shown in Figures 2 and 10, human‑composed music gravitates toward I‑IV repetitions (I‑IV‑I and IV‑I‑IV are Lastfm’s two most frequent trigrams), I‑V patterns (I‑V‑I, V‑I‑V), and the functional IV‑I‑V and IV‑V‑I, with the minor‑mode oscillation i‑bVI‑i close behind. Udio follows the same profile and keeps the Aeolian idioms bVI‑bVII‑i, i‑bVI‑bVII, and bVII‑i‑bVI among its most frequent trigrams. Progressions of this type, particularly the bVI‑bVII‑i pattern, have been identified as characteristic harmonic idioms of rock music, functioning as Aeolian cadences with subdominant–dominant–tonic implications distinct from conventional major‑minor tonality (Biamonte, 2010). Suno’s top ranks are different in kind: even with repetition patterns counted, its four most frequent trigrams are the four three‑chord windows of the Axis loop (IV‑I‑V, vi‑IV‑I, I‑V‑vi, V‑vi‑IV).

Figure 10

6.3 Structural similarities and differences

The n‑gram analysis of section‑label sequences and the count of distinct full song forms provide complementary views of formal diversity across the three collections (Figure 7).

Suno exhibits the strongest convergence toward simple, repetitive structures. Verse–chorus cycles dominate its most frequent section‑label tetragrams, and its tracks collapse to the fewest distinct full song forms of the three collections ($3,349$, against $3,814$ for Udio and $6,020$ for Lastfm): the human collection is the most formally diverse, and Udio sits closer to it than Suno does. The section‑label tetragram heatmap (Figure 7) confirms that Udio’s tetragram frequencies are closer to Lastfm’s than Suno’s are ($r=0.954$ versus $0.913$). The clearest Suno‑specific pattern is the closing formula chorus–verse–chorus–outro, its second most frequent tetragram: with $4,032$ occurrences against $2,061$ in Lastfm and $913$ in Udio, Suno closes a verse‑chorus cycle directly with an outro at more than twice Lastfm’s per‑song rate and three times Udio’s. Instrumental sections as connectors inside verse‑chorus cycles are, by contrast, not an AI signature: verse‑chorus‑inst‑verse appears at essentially the same per‑song rate in all three collections (Lastfm $2,618$; Suno $2,103$; Udio $1,599$).

The normalized musical form trigram profiles per genre (Figure 11) reveal a consistent pattern across styles. Both Suno and Udio show a reduced range of formal options relative to Lastfm, with intro‑verse‑chorus sequences markedly under‑represented; solo sections are under‑represented in Suno, at roughly half Lastfm’s rate, while Udio stays closer to the human level. This is consistent with how both platforms operate: users generate a song segment that begins directly with a verse‑chorus cycle and can later be extended via the web interface. A plausible result is that intro sections, which precede the core structure, are rarely generated in the initial output, before any user‑driven extension. The same genre‑ontology caveat raised for Figure 5 applies here: the Jazz polygon in Figure 11 should not be read as a direct cross‑collection comparison of jazz formal conventions, since the tracks grouped under that label differ substantially in character across the three collections.

Figure 11

A key confound is song duration. Suno and Udio tracks average $175$ and $143$ s respectively, against $261$ s for Lastfm (Figure 12). Shorter tracks have fewer opportunities for structural elaboration (bridges, instrumental breaks, key changes), so the reduced diversity observed in Suno and Udio may partly reflect a generation‑length constraint rather than a purely compositional bias. Udio’s duration distribution is notably bimodal: a sharp spike below one minute reflects tracks published at the platform’s default clip length of 32 s without extension, while the secondary peak near $130$ s corresponds to tracks generated at the platform’s longest available clip length.

Figure 12

A further caveat concerns formal labels and musical function. The All‑In‑One model assigns labels based on acoustic properties, but a verse–chorus alternation in a generated track may represent static loop repetition rather than the narrative‑arc form those labels imply in conventional songwriting (Nieto et al., 2020). Distinguishing conventional pop form from non‑teleological loop structures would require additional features such as melodic contour, rhythmic analysis, or lyric structure.

6.4 Outcome homogenization and creative agency

The harmonic compression documented in Section 6.2 and the structural narrowing documented in Section 6.3 both point toward a form of outcome homogenization, consistent with the concept of algorithmic monoculture, where shared components across AI systems increase the likelihood of convergent outputs (Bommasani et al., 2022). Suno exemplifies this pattern, converging on a statistical peak of canonical pop progressions (Figure 6, Table 4), also visible in the most common harmonic tetragrams (Figure 10). The output does not reflect the variety of harmonic practice or formal organization found in human‑composed music. Several mechanisms may contribute to this convergence. First, training data bias amplification, in which generative models learn to reproduce only the most frequent patterns in their training corpora. Since contemporary pop music dominates streaming platforms and likely constitutes a significant portion of Suno’s training data, these harmonic patterns are overrepresented in its outputs. Second, optimization toward perceptual plausibility encourages models to generate music that sounds ‘familiar’ to broad audiences, reinforcing existing harmonic conventions rather than exploring innovative progressions. Third, the lack of harmonic‑prompting capabilities in the platforms at the time of data collection (2024) means that users could not directly specify harmonic content, instead relying on the models’ learned biases toward common progressions. Fourth, Suno’s developers may intentionally be targeting users focusing on commercial modern pop styles. Our data cannot determine which of these factors is primary, but the contrast with Udio is informative. Udio seems to produce more diverse harmonic and structural outputs than would be expected from a presumably comparable training distribution, suggesting that model architecture and generation strategy play a role beyond the training data alone.

A restricted harmonic vocabulary is not in itself a compositional deficiency: 12‑bar blues, punk, and minimal techno are all built on narrow chord sets by deliberate stylistic choice. The data suggest that Suno applies a single harmonic template across all stated genres, a platform‑level convergence that is independent of user intent. This is consistent with Suno’s own positioning: the platform’s leadership has described it explicitly as gamifying musical creation.¹² This raises questions about whether these tools genuinely support creative agency among musicians with any degree of musical knowledge (Lin et al., 2023; Tan, 2024; Wiggins, 2007, 2016). The prevalence of simple, repetitive formal structures (as revealed by our section‑label analysis shown in Figure 7) further suggests that the platforms may be optimized to generate familiar, commercially viable content rather than to support experimentation or skill acquisition. This optimization strategy may serve commercial interests, but it limits the creative potential of these platforms as compositional tools.