Skip to main content
Have a personal or library account? Click to login
Data‑Driven Analysis of Text‑Conditioning in AI‑Generated Music: A Case Study with Suno and Udio Cover

Data‑Driven Analysis of Text‑Conditioning in AI‑Generated Music: A Case Study with Suno and Udio

Open Access
|May 2026

Figures & Tables

Figure 1

Suno’s interface in ‘Custom’ mode. The user can manually input lyrics and tags (style of the music). Last accessed December 16, 2024.

Figure 2

Udio’s interface (as of February 18, 2025). Custom mode allows for manual adjustments even if the prompts are automatically generated.

Table 1

The subset of the metadata from Suno and Udio we analyzed.

FieldSunoUdio
idUnique song IDUnique song ID
titleTitle of the songTitle of the song
tagsTags derived from the user inputUser tags
replaced_tagsN/ADictionary with tags replacements
lyricsN/ASong lyrics
promptSong lyricsUser input
gpt_description_promptUser prompt for lyrics generationN/A
optimized_promptN/ARefined user input
Table 2

The 15 most popular languages for lyrics in the dataset. The percentage refers to the prevalence in each platform.

LanguageISO 639‑3UdioSuno
Englisheng_Latn71.39%46.75%
Germandeu_Latn3.68%8.87%
Russianrus_Cyrl2.99%6.68%
Spanishspa_Latn3.28%4.58%
Portuguesepor_Latn1.68%3.55%
Koreankor_Hang3.21%3.00%
Chineseyue_Hant1.77%3.33%
Indianind_Latn0.27%3.26%
Frenchfra_Latn1.81%2.15%
Japanesejpn_Jpan1.45%1.92%
Turkishtur_Latn0.29%1.66%
Italianita_Latn1.06%1.29%
Thaitha_Thai0.05%1.26%
Vietnamesevie_Latn0.09%1.11%
Polishpol_Latn0.77%0.94%
TOTAL93.79%90.35%
Figure 3

Clustering of prompts embeddings. The names for each cluster are manually defined after checking their content.

Table 3

Macro‑categories (manually defined) and HDBSCAN clusters (manually renamed) in the lyrics‑embedding space with their respective sizes.

CategoryClusters
abstract (1719)afrofuturism (26), afrofuturism (64), carpe diem (70), chaos (21), clashing (60), dreams (49), dreams (31), flying (199), spirituality (42), mask (24), mirrors (53), money (26), photo (35), post‑atomic (32), religion (854), shadows (48), tired (37), war (48)
animals (708)animals (74), bees (38), birds (40), butterfly (23), capybara (33), cats (280), dogs (169), fireflies (27), frog (24)
celebration (232)birthday (126), halloween (31), xmas (75)
daily life (345)chores (23), daily work (68), monday (54), monotony (25), rent (37), school (109), weekend (29)
dance (515)beat (31), beat (64), last (49), night (31), night (125), groove (125), heartbeat (29), moonlight (25), swing (36)
driving (210)driving (83), road trip (47), speed (80)
family (241)family (52), father (43), friends (87), mother (59),
fantasy (893)demons (21), fantasy (562), shadows (24), spooky (117), vampire (29), vikings (83), werewolf (57)
feelings (809)break free (109), fade away (26), good ol days (31), happiness (24), loneliness (131), madness (116), old place (46), pain (30), peaceful (28), revenge (23), run free (25), runaway (29), tears (22), weariness (114), yesterday (55)
food (784)candy (60), cheese (522), coffee (101), fruit (101)
genre (817)blues (60), country music (301), emo (195), guitar (52), heavy metal (29), rock ‘n’ roll (50), trap‑like (130)
location (679)america (38), australia (21), beach (28), capitals (23), desert (58), earth (96), egypt (35), forest (65), river (26), sea (289)
love (2182)apology (21), goodbye (85), heartbrake (172), i miss you (45), always (74), breakup (81), burning (37), can’t wait (52), crazy (29), distance (50), dream (31), electric (40), eyes (97), feel (41), forever (26), forever (406), hand (20), holding on (33), letting go (22), longing (80), loss (28), missing (27), stay (37), time (26), togetherness (41), unrequited (292), wait (33), wandering (38), whisper (76), whisper (47), return (95)
meme (412)fck (183), memes (91), pp (118), weed (20)
mixed language (769)chinese (93), hindi (23), indonesian (24), jamaican (86), japanese (355), korean (108), russian (52), spanish (28)
motivational (288)new dawn (50), dreams (40), phoenix (59), rising (61), shine (21), unstoppable (57)
other (557)alphabet (27), boots‑pub? (27), gpt glitch? (25), absurd? (25), poe‑raven (39), short+instr. (414)
politics (143)palestine (29), protest songs (71), trump+biden (43)
sports (254)sports (197), training (57)
stars/night (648)cosmic (460), quiet night (39), stars (32), nightsky (117)
technology (899)AI (663), code (110), crypto (40), digital (55), math (31)
time (336)midnight (32), midnight (130), midnight+love (22), morning (38), sunset (84), time (30)
urban (595)city (135), city (165), lost (22), neons (233), street (40)
videogames (126)helldivers (20), pokemon (42), videogames (64)
weather/seasons (949)autumn (51), frozen (39), moonlight (140), rain (62), rain+dancing (183), rain+love (49), rainy day (39), summer (211), sunshine (155), sunshine+love (20)
outliers (25794)outliers (25794)
Figure 4

Clusters of lyrics (from both Suno and Udio) obtained from the HDBSCAN algorithm applied to a five‑dimensional UMAP reduction of the embedding space. The scatterplot is then created on a two‑dimensional reduction of the same space. Colors represent macro‑categories and text annotations refer to the specific clusters, as shown in Table 3.

Figure 5

Word cloud for Suno (left) and Udio (right). Font size is scaled according to prevalence.

Table 4

Manually created high‑level taxonomy derived from tags used more than 10 times. For Suno and Udio separately we indicate the expected number of tags λ and the probability of seeing more than one in a string according to a fitted Poisson distribution. UNDEFINED refers to tags that appear in a string but don’t match with our vocabulary.

Categoryn. of tagsλsunoλudioPsuno(X>1)Pudio(X>1)
GENRE/STYLE6571.15e + 003.90e + 006.84e‑019.80e‑01
QUALIFIER3248.38e‑013.56e+005.67e‑019.72e‑01
INSTRUMENT1082.54e‑012.58e‑012.24e‑012.28e‑01
STRUCTURE681.34e‑012.96e‑011.26e‑012.56e‑01
VOICE511.27e‑016.08e‑011.19e‑014.55e‑01
YEAR222.98e‑025.39e‑022.93e‑025.25e‑02
KEY103.65e‑033.53e‑033.64e‑033.53e‑03
BPM61.63e‑033.93e‑041.63e‑033.93e‑04
UNDEFINED−1.26e+001.15e+007.17e‑016.84e‑01
Figure 6

Clusters of the most common tags (combined ranking from both services). Colors correspond to macro‑categories defined manually. Text corresponds to the most prevalent tag in each cluster of the clusters we find with HDBSCAN. Grey circles indicate outliers.

Table 5

The 10 most replaced artists found in Udio’s metadata under replaced_tags. The full list contains 703 artists.

Artist#Replaced Tags
XXXTentacion26emo rap, alternative r&b, hip hop, contemporary r&b, r&b, pop rap, aggressive, self‑hatred, boastful, depressive
Drake19male vocalist, pop rap, contemporary r&b, hip hop, r&b, alternative r&b, atmospheric, introspective, apathetic, mellow, bittersweet
Taylor Swift18alt‑pop, singer–songwriter, synthpop, nocturnal, romantic, love, atmospheric, lonely, sentimental, longing, concept album, lethargic, passionate, 2020s
Foo
Fighters
18male vocalist, alternative rock, post‑grunge, acoustic rock, energetic, melodic
The
Beatles
17male vocalist, psychedelic pop, pop rock, psychedelia, sunshine pop, art pop, melodic, lush, love, fantasy, optimistic, dense, pastoral
Depeche Mode17male vocalist, synthpop, downtempo, ambient pop, electronic, melancholic, melodic, calm, soothing, lush, mellow, nocturnal
Adele17female vocalist, pop soul, adult contemporary, pop, blue‑eyed soul, passionate, sad, sentimental
J. S. Bach16classical music, baroque music
The Weeknd14male vocalist, alternative r&b, electropop, r&b, electronic, synthpop, nu‑disco, party, hedonistic
ABBA14female vocalist, europop, euro‑disco, dance, pop, disco, optimistic, energetic, uplifting, melodic, rhythmic, party, lush
Table 6

Top 25 most prevalent metatags. In cases where numbers appear, e.g., Chorus 2, they were stripped and merged into one category.

SequenceSunoUdioTotal
verse737221454888270
chorus486701667265342
bridge17826466522491
outro616323868549
pre‑chorus372529146639
end37472894036
intro243611713607
instrumental26827103392
drop66214032065
guitar solo12128352047
hook9124681380
break9432121155
interlude538398936
fade out644208852
instrumental break504280784
solo537236773
instrumental solo63339672
instrumental intro59664660
breakdown260277537
refrain314203517
instrumental interlude45255507
yeah3466469
pre‑hook42615441
build163267430
final chorus144255399
Table 7

A few examples of metatags that mention real artists.

ArtistMetatag
Bob Marleyproduced by bob marley and lee perry
Journeyjourney separate ways synth arpeggio
Kanye Westverse 2: kanye west
Madonnainfluence: madonna, michael jackson
DOI: https://doi.org/10.5334/tismir.273 | Journal eISSN: 2514-3298
Language: English
Page range: 194 - 209
Submitted on: Apr 30, 2025
Accepted on: Apr 3, 2026
Published on: May 7, 2026
Published by: Ubiquity Press
In partnership with: Paradigm Publishing Services
Publication frequency: 1 issue per year

© 2026 Luca Casini, Laura Cros Vila, David Dalmazzo, Anna-Kaisa Kaila, Bob L.T. Sturm, published by Ubiquity Press
This work is licensed under the Creative Commons Attribution 4.0 License.