Listeners identified the emotion a game music generator intended only 34% of the time, against 25% by chance, and the human-composed pieces in the same test scored the same 34% (Ferreira, Lelis and Whitehead, AIIDE 2020). That result frames the field: the target is hard even for human music. Hutchings and McCormack’s adaptive system won a significant within-group immersion result, yet none of its four between-group comparisons fell below p = 0.05 (IEEE Transactions on Games 2019). On Steam, games disclosing generative AI draw a 68.4% recommendation rate against 86.3% for procedural-content games, though that figure covers whole games, not music (Bazzaz and Cooper, arXiv preprint 2026). We aggregated data from IEEE Transactions on Games, AIIDE, ISMIR, the IEEE Conference on Games, and peer-reviewed player studies. This roundup stays on systems and their evaluations, so read it alongside our game soundtrack statistics for the commercial side of game scores.
TL;DR
- 34% correct on valence plus arousal vs 25% by chance, with human pieces also at 34% (Ferreira et al., AIIDE 2020, 116 participants).
- Immersion p = 0.121 for Zelda and p = 0.053 for StarCraft II, adaptive system vs original score (Hutchings and McCormack, IEEE Transactions on Games 2019, 34 players).
- Adaptive music raised presence and emotional impact in VR archery (Wen et al., IEEE TVCG 2026, 22 participants).
- Overall quality on a 1 to 5 scale: human 4.11, best generator 3.94 (Ferreira et al., AIIDE 2022, 199 valid answers).
- Audio-only game agent win ratio rose from 0.52 to 0.94 with adaptive music (Khan et al., IEEE CoG 2023, 90 rounds per condition).
- MusicGen produced 8-second clips in less than 4 seconds on one test machine (Marrinan et al., arXiv preprint 2024).
- 5,278 songs from 397 games in the NES Music Database (Donahue et al., ISMIR 2018).
- 98,940 gameplay clips paired with music from 389 games (Cardoso et al., FDG 2024).
- 38.87% of interview codes were concerns and 13.09% were benefits (Worrall and Collins, IEEE Transactions on Games 2023, 11 composers).
- Steam recommendation rate of 86.3% for procedural-content games vs 68.4% for generative AI games (Bazzaz and Cooper, arXiv preprint 2026).
1. Player Studies: Tension Moves, Immersion Is Harder to Prove
None of the four between-group comparisons in the Hutchings and McCormack study fell below p = 0.05, even though the paper’s within-group test did. The headline “players felt more immersed” is true of how each person shifted, not of a clean win over the original soundtrack. Samples this small can only detect large effects, so a near miss at p = 0.053 is evidence of neither success nor failure.
Tension is the sturdier result. In Galactic Escape, the Simon Fraser University test game, an adaptive score raised experienced tension even when it was mapped backwards to the game’s tension curve. That suggests music that audibly responds may matter more than music that responds correctly, which should give pause to anyone spending weeks on mapping rules. Wen et al. (2026) add a warning from VR: moderate musical tension improved aim and aiming time, high tension impaired both, an inverted-U shape.
| Metric | Value | Source |
|---|---|---|
| Fixed soundtrack vs adaptive generative soundtrack in VR archery, within-subjects design | 22 participants, 4 tension phases; presence and emotional impact rose | Wen et al., IEEE Transactions on Visualization and Computer Graphics 2026 |
| Immersion, adaptive system vs original score of the game Zelda: MoS | p = 0.121, η² = 0.042 (34 players, 2 groups of 17, 20-minute sessions) | Hutchings and McCormack, IEEE Transactions on Games 2019 |
| Immersion, adaptive system vs original StarCraft II score | p = 0.053, η² = 0.079 | Hutchings and McCormack, IEEE Transactions on Games 2019 |
| Reported effect of music on immersion, within each group, with the adaptive system | Significant increase in both groups (p<0.05, Wilcoxon signed-rank test) | Hutchings and McCormack, IEEE Transactions on Games 2019 |
| Experienced tension, mean by soundtrack: tension-mapped 0.62, inverse-mapped 0.61, no music 0.55, neutral 0.49 | F(3, 87) = 3.662, p = .015, partial η² = .112 (30 players kept of 35 recruited) | Plut, Simon Fraser University PhD thesis 2022 |
| Generative score vs composed linear score vs composed adaptive score | 48 participants; linear rated highest on all questions, generative a close second on emotional congruency, immersion and preference | Plut, Simon Fraser University PhD thesis 2022 |
| Generative vs linear soundtrack in a Tetris-like game | 39 undergraduates; hypotheses not supported, players could not tell the two apart, flow was still affected | Sites and Potter, Game Studies 2018 |
| Mood-rated training set of a web game music generator, then A/B tests against static music | 80 rated pieces; adaptiveness did not necessarily improve on composed music | Adam, Cal Poly MS thesis 2014 |
The two oldest rows add nuance. Sites and Potter saw higher flow (a state of focused, rewarding absorption) among those who played the generative version first (sample from the Tampere University Playlab summary), and the AUD.js thesis found no automatic gain over a composed score.
Most recent available data for this theme: Wen et al. 2026. Every other study here dates from 2014 to 2022.
2. Listener Evaluations: Does the Intended Emotion Survive?
Human-composed pieces scored no better than the generator in the Bardo Composer test: both landed on 34%. That is a reality check on the task more than a failure of the model. Identifying valence (how pleasant or unpleasant music feels) and arousal (how calm or energetic it feels) at once is simply hard, and any vendor claiming near-perfect emotional targeting should be asked what the human baseline scored.
Quality ratings tell a tighter story. The 2022 AIIDE study put its best generator within 0.17 points of human music (4.11 minus 3.94), though only 199 answers survived filtering out of 1,600 recruits, a reminder that crowd-worker listening data is noisy before it is informative. A smaller test by Yan (arXiv preprint 2023) on 16 rhythm game piano arrangements also found listeners preferred human-made music; it gives no respondent count.
| Metric | Value | Source |
|---|---|---|
| Bardo Composer (a generator for tabletop RPG sessions): correct on valence and arousal together | 34% vs 25% by chance, p = 0.02 (116 participants, 2 groups of 58) | Ferreira, Lelis and Whitehead, AIIDE 2020 |
| Bardo Composer: correct on intended arousal, best reported part (episode 4, part 1) | 87% of participants | Ferreira, Lelis and Whitehead, AIIDE 2020 |
| Overall quality, 1 to 5 scale, mean ± SD | Human 4.11 ± 0.73, tree-search method (PUCT) 3.94 ± 0.76, CS 3.68 ± 0.96, SBBS 3.62 ± 1.01 (199 valid answers of 1,600 recruited, 80 pieces per generator) | Ferreira et al., AIIDE 2022 |
| Composed emotional clips, listener agreement in the one-rank design | 58.5% overall: valence 59.2%, arousal 59.4%, tension 56.8% (180 participants, 90 clips) | Plut, Simon Fraser University PhD thesis 2022 |
| Sets of clips where listeners confirmed the composer’s labels | 19 of 30 matched, 10 disagreed, 1 had no clear agreement | Plut, Simon Fraser University PhD thesis 2022 |
Even clips written to a target emotion missed it: 10 of 30 sets went the other way. Generative systems trained on such labels inherit that disagreement.
Most recent available data: Ferreira et al., AIIDE 2022. We found no 2025 or 2026 listener study of this size for game music.
3. System Measurements: Speed, Stability and Behavior
An audio-only game agent’s win ratio climbed from 0.52 to 0.94 once background music tracked player health and distance. No human played in that IEEE CoG study, and it uses simple volume rules over separated instrument groups. It still makes a useful point: adaptive music can carry game state, not only mood. Rhythm games take that idea furthest, though their tracks are licensed, not generated, as our rhythm game music licensing statistics show.
Speed is uneven. A clip model beat real time on one test machine, while a heavy transformer pipeline still ran offline. Prompt stability is the newer concern: Wu and Si (IVA 2026) tested whether small prompt edits cause abrupt musical jumps, which would break a score that changes prompts during play.
| Metric | Value | Source |
|---|---|---|
| Time to generate an 8-second MusicGen clip (Threadripper 2950X CPU, RTX 2080 GPU) | Less than 4 seconds | Marrinan et al., arXiv preprint 2024 |
| Update cycle of an adaptive music system’s game-state messages | Every 30 ms, with agents writing two measures (bars of music) at a time | Hutchings and McCormack, IEEE Transactions on Games 2019 |
| Unique arrangements from one generative adaptive score | Almost 14 trillion, generated offline because the MMM model is too heavy for real time | Plut, Simon Fraser University PhD thesis 2022 |
| Audio similarity after a small prompt change, MusicGen-large, CLAP cosine similarity (1.0 = identical) | 0.7666 with one word swapped, 0.8202 with an intensity shift (75 prompt groups across all conditions, 6 seeds) | Wu and Si, IVA 2026 |
| Win ratio, audio-only agent with FFT audio encoder, original vs adaptive music | 0.52 vs 0.94; average health-point lead 27.03 vs 136.14 (90 rounds per condition) | Khan et al., IEEE CoG 2023 |
| Win ratio, audio-only agent with mel-spectrogram encoder, original vs adaptive music | 0.56 vs 0.68 (90 rounds per condition) | Khan et al., IEEE CoG 2023 |
| Frechet Audio Distance to a 32-hour cinematic reference set, tabletop RPG campaign, lower is closer | Best generator 3.00 vs original human soundtrack 9.66 | Marra and Ferreira, LAMIR workshop (arXiv 2024) |
| Story fit: divergence from the original background music (KLD), lower is closer | Emotion-prompt version 3.34 ± 1.89 vs baseline 4.84 ± 2.98 | Marra and Ferreira, LAMIR workshop (arXiv 2024) |
The Frechet row needs care. A lower distance means audio statistically closer to a cinematic reference set, not better music than the human soundtrack. The Babel Bardo paper used automatic metrics across 2 campaigns, with no human preference test, and treats listener studies as future work.
4. Training Data: Corpora Built on Old Sound Chips
The NES Music Database holds 5,278 songs from 397 games, about 46.1 hours in total. A 1980s console remains central to game music research because its songs are compact, multi-instrumental and symbolic, not because it resembles a modern orchestral score. The gap between training data and shipped soundtracks is worth keeping in mind: for licensed stock alternatives, see our production music library statistics.
The evaluation habit matters as much as the size. Neither the NES Video-Music Database nor YM2413-MDB baseline used a listening test, leaning on objective metrics and classifiers. Conditioned generation got closer to human music on all five reported metrics, and yet a genre classifier still recognized generated pieces less often than human ones.
| Metric | Value | Source |
|---|---|---|
| NES Music Database size | 5,278 songs, 397 games, 296 composers | Donahue et al., ISMIR 2018 |
| NES Music Database volume | 2,325,636 notes, 46.1 hours | Donahue et al., ISMIR 2018 |
| NES Video-Music Database size | 98,940 clips of 15 seconds from 389 games, 474 hours of video | Cardoso et al., FDG 2024 |
| Pieces with paired gameplay clips, NES Video-Music Database | 4,070 of the 5,278 pieces | Cardoso et al., FDG 2024 |
| Grooving pattern similarity: model conditioned on video, unconditioned model, human pieces | 0.821, 0.694, 0.999 (140 pieces per set) | Cardoso et al., FDG 2024 |
| Genre classifier accuracy on generated pieces vs human pieces | 22% vs 34% | Cardoso et al., FDG 2024 |
| YM2413-MDB, 1980s FM chip game music with emotion tags | 669 files, 15.4 hours, 19 emotion tags, 2 annotators plus 3 verifiers per song | Choi et al., ISMIR 2022 |
| Four-quadrant emotion recognition accuracy, YM2413-MDB | 0.65 best on audio (short-chunk ResNet), 0.48 best on symbolic music (logistic regression) | Choi et al., ISMIR 2022 |
Emotion labels remain the scarcest resource. VGMIDI began with 823 pieces of which 95 were annotated, yielding 966 labelled phrases and a sentiment classifier at 89.83 ± 3.14% (Ferreira and Whitehead, ISMIR 2019). Its labelled set later grew to 200 pieces (Ferreira, Lelis and Whitehead, AIIDE 2020). Structure labels are thinner still: a 2026 preprint by Luo and Reiss built a game audio dataset with 309 structural annotations, arguing that weak grasp of structure is one reason the industry rarely uses neural generators.
Most recent available data: Donahue et al. 2018, Ferreira and Whitehead 2019, Choi et al. 2022.
5. Composers: Quality and Identity Outrank Job Fear
‘AI does not generate original, human quality music’ was the largest sub-theme in one interview study of professional game composers, with 191 of 1,199 codes. Job security drew just 11. The instinct to read this field as a labor panic does not match what these 11 composers actually spent their interview time on: quality, creative control and ownership. That reframes the design brief for tools, and it connects to the legal questions tracked in our AI music copyright claims statistics.
These are code counts from a reflexive thematic analysis by one researcher, so they measure how much the topic came up, not how many people hold a view. Treat them as a map of the conversation, not a poll.
| Metric | Value | Source |
|---|---|---|
| Study sample, professional game composers with at least one published game | 11 people, about 13 hours of interviews, 1,199 codes | Worrall and Collins, IEEE Transactions on Games 2023 |
| Share of codes about concerns over AI music technology | 38.87% | Worrall and Collins, IEEE Transactions on Games 2023 |
| Share of codes on the clash between creative ego and AI vs share on benefits of AI | 29.36% vs 13.09% | Worrall and Collins, IEEE Transactions on Games 2023 |
| Codes on ‘music AI clashes with the creative ego’ vs ‘AI can benefit workflow’ | 133 vs 70 | Worrall and Collins, IEEE Transactions on Games 2023 |
| Codes on concerns about job security vs concerns about musical representation and ownership | 11 vs 83 | Worrall and Collins, IEEE Transactions on Games 2023 |
6. Field Context: Few Systems Studied, Wary Steam Players
The most complete academic census we could read, from 2022, covers just 34 generative music systems from games, 19 of them from the prior decade. A newer overview by Luo and Reiss (AES 2025) says academic prototypes and commercial use still diverge because the two sides prioritize novelty, reliability and resources differently. No public dataset we found counts shipped games that use procedural music, so any such figure online deserves suspicion.
The nearest proxy for player sentiment comes from Steam, and it covers whole games, not soundtracks. This 2026 preprint used procedural content generation as a baseline for games disclosing generative AI. Reviewers tended to read generative AI as a sign of low developer investment, which is a warning for any music tool marketed on that label.
| Metric | Value | Source |
|---|---|---|
| Generative music systems from games in one academic survey | 34 systems, 19 from the 10 years before the survey | Plut, Simon Fraser University PhD thesis 2022 |
| Share of Steam releases disclosing generative AI | 5.6% in 2024, 10.3% in 2025 | Bazzaz and Cooper, arXiv preprint 2026 |
| Recommendation rate, procedural-content games vs generative AI games | 86.3% vs 68.4% (5,186 vs 5,970 games, 508,192 English reviews) | Bazzaz and Cooper, arXiv preprint 2026 |
| Logistic regression result for generative AI games vs procedural-content games | Odds ratio 0.63 (95% CI [0.62, 0.64], p < .001) | Bazzaz and Cooper, arXiv preprint 2026 |
| Recommendation rate in free-to-play games vs paid games, procedural vs generative AI | Free-to-play 83.1% vs 50.1%; paid 86.9% vs 78.9% | Bazzaz and Cooper, arXiv preprint 2026 |
| Recommendation rate in reviews that show awareness of generative AI | 49.5% (63.7% negative sentiment, n = 7,834) | Bazzaz and Cooper, arXiv preprint 2026 |
Summary: Procedural and Generative Game Music Systems by the Numbers
| Metric | Value | Source |
|---|---|---|
| Bardo Composer listeners correct on valence and arousal | 34% vs 25% by chance (116 participants) | Ferreira, Lelis and Whitehead, AIIDE 2020 |
| Overall quality, human vs best generator, 1 to 5 scale | 4.11 vs 3.94 (199 valid answers) | Ferreira et al., AIIDE 2022 |
| Immersion, adaptive system vs original StarCraft II score | p = 0.053 (34 players) | Hutchings and McCormack, IEEE Transactions on Games 2019 |
| VR archery, fixed vs adaptive generative music | 22 participants, presence and emotional impact rose | Wen et al., IEEE Transactions on Visualization and Computer Graphics 2026 |
| Generative vs linear soundtrack in a Tetris-like game | 39 undergraduates, hypotheses not supported | Sites and Potter, Game Studies 2018 |
| Unique arrangements from one generative adaptive score | Almost 14 trillion, offline only | Plut, Simon Fraser University PhD thesis 2022 |
| MusicGen 8-second clip generation time | Less than 4 seconds | Marrinan et al., arXiv preprint 2024 |
| MusicGen-large similarity after an intensity shift in the prompt | 0.8202 CLAP cosine similarity | Wu and Si, IVA 2026 |
| Audio-only agent win ratio, original vs adaptive music | 0.52 vs 0.94 | Khan et al., IEEE CoG 2023 |
| Frechet Audio Distance, best generator vs human soundtrack | 3.00 vs 9.66 | Marra and Ferreira, LAMIR workshop (arXiv 2024) |
| NES Music Database | 5,278 songs, 397 games, 46.1 hours | Donahue et al., ISMIR 2018 |
| NES Video-Music Database | 98,940 clips, 389 games | Cardoso et al., FDG 2024 |
| YM2413-MDB | 669 files, 15.4 hours | Choi et al., ISMIR 2022 |
| Composer codes on ‘not original, human quality’ vs job security | 191 vs 11 of 1,199 codes | Worrall and Collins, IEEE Transactions on Games 2023 |
| Steam recommendation rate, procedural vs generative AI games | 86.3% vs 68.4% | Bazzaz and Cooper, arXiv preprint 2026 |
Methodology and Sources
We read each study or its repository page and report samples as printed. No figure comes from an SEO blog or a listicle.
- Wen et al., Tuning Immersion and Performance with Adaptive Generative Music in VR, IEEE TVCG 2026
- Hutchings and McCormack, Adaptive Music Composition for Games, IEEE Transactions on Games 2019
- Plut, Application and evaluation of affective adaptive generative music for video games, Simon Fraser University 2022
- Ferreira, Lelis and Whitehead, Computer-Generated Music for Tabletop Role-Playing Games, AIIDE 2020
- Ferreira et al., Controlling Perceived Emotion in Symbolic Music Generation with Monte Carlo Tree Search, AIIDE 2022
- Ferreira and Whitehead, Learning to Generate Music With Sentiment, ISMIR 2019
- Worrall and Collins, Considerations and Concerns of Professional Game Composers Regarding Artificially Intelligent Music Technology, IEEE Transactions on Games 2023
- Donahue et al., The NES Music Database, ISMIR 2018
- Cardoso et al., The NES Video-Music Database, FDG 2024
- Choi et al., YM2413-MDB, ISMIR 2022
- Khan et al., Fighting Game Adaptive Background Music for Improved Gameplay, updated from IEEE CoG 2023
- Wu and Si, Evaluating Prompt Robustness in Text-to-Audio Systems, IVA 2026
- Marrinan et al., AI audio for user-generated game content, arXiv preprint 2024
- Marra and Ferreira, Long-Form Text-to-Music Generation with Adaptive Prompts, LAMIR workshop
- Bazzaz and Cooper, Player Perceptions of Generative AI in Games, arXiv preprint 2026
- Luo and Reiss, Procedural Music Generation Systems in Games, AES 2025 and Supervised Learning for Game Music Segmentation, arXiv preprint 2026
- Sites and Potter, Everything Merges with the Game, Game Studies 2018, with the sample taken from the Tampere University Playlab summary
- Adam, Procedural Music Generation and Adaptation Based on Game State, Cal Poly MS thesis 2014 and Yan, Generating Rhythm Game Music with Jukebox, arXiv preprint 2023
- Data watch: Most samples here are small, so a non-significant result is not proof of no effect. Worrall and Collins state 29 sub-themes in one place and 30 in another (their table lists 29), and their count of AAA composers does not reconcile, so we quote neither. Plut’s PreGLAM chapter says 50 participants in its introduction and 48 in its methods; we use 48. In Wu and Si, the full-text table we read shows lower Stable Audio similarity under rephrasing than the abstract’s ranking implies, so we quote only the MusicGen-large figures, which agree with the abstract. Khan et al. used no human players and report no significance tests. The Steam preprint covers whole games and is not peer reviewed, and its median playtime differs between a table and the text, so we omit that figure. The composer interviews date from 2023 and the systems census from 2022, so both are due for a successor. Vendor-run adaptive music surveys that circulate online could not be traced to a primary document and were dropped. Commercial adoption counts were not available in a form we could verify.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when the AIIDE 2026 and ISMIR 2026 proceedings are published.