treadon/banger-scorer-generated-songs

Banger Scorer Generated Songs

230 AI-generated songs across 10 genres and 5 languages, all scored by the banger scorer. Includes 10 genre tests (20 songs each) plus 1 banger-optimized run (30 songs) that used data-driven parameter selection to maximize scores.

Every song includes its MP3 audio, generation metadata (BPM, key, seed, caption prompt), and banger score. Useful for research into AI music quality, training better scorers, or just listening to what works and what does not.

Dataset Description

What is this?

A systematic evaluation of AI music generation quality. 230 songs were generated with ACE-Step 1.5 on an Apple M4 Pro, spanning a grid of caption styles, BPMs, keys, and languages. Each song was then scored by the banger scorer (MERT + MLP, trained on FMA play counts).

The dataset captures what an automated quality model thinks about AI-generated music -- and reveals clear patterns in what makes some generated songs score higher than others.

Structure

230 songs total:
  ├── test01/  (20 songs) -- Hip Hop (English)
  ├── test02/  (20 songs) -- Pop/Dance (English)
  ├── test03/  (20 songs) -- R&B/Soul (English)
  ├── test04/  (20 songs) -- Latin/Reggaeton (Spanish)
  ├── test05/  (20 songs) -- Bollywood (Hindi)
  ├── test06/  (20 songs) -- Punjabi/Bhangra (Punjabi)
  ├── test07/  (20 songs) -- C-Pop (Chinese)
  ├── test08/  (20 songs) -- Rock/Alternative (English)
  ├── test09/  (20 songs) -- Electronic/EDM (Instrumental)
  ├── test10/  (20 songs) -- Acoustic/Folk (English)
  └── bangers/ (30 songs) -- Banger-optimized run (mixed EDM/Bhangra/Bollywood)

Per-Song Fields

Field	Type	Description
audio	MP3 (192kbps)	The generated song, 2 minutes, 48kHz stereo
genre	string	Genre category (e.g., "EDM", "Hip Hop", "Bollywood")
bpm	int	Beats per minute used for generation
key	string	Musical key used for generation (e.g., "Eb minor", "D major")
seed	int	Random seed for deterministic reproduction
caption	string	Full text prompt given to ACE-Step
banger_score	float (0-10)	Score from the banger scorer model
language	string	Lyrics language (en/es/hi/pa/zh or instrumental)

Generation Details

Model

ACE-Step 1.5 -- a full-song AI music generator with vocals, instruments, and structure. Generates ~2-minute songs from a text caption and optional lyrics.

Hardware

Apple M4 Pro with 64 GB unified memory. Each song took ~110 seconds to generate (language model planning ~27s, DiT diffusion ~9s for 8 turbo steps, VAE decoding ~18s, plus overhead). Total generation time for all 230 songs: ~7 hours. Scoring all 230 songs took ~8 minutes.

200/200 success rate on genre tests, 30/30 on the banger run. Zero generation failures. Achieved by spawning a fresh subprocess per song to avoid MPS memory accumulation.

Parameter Grid

Each genre test sampled 20 combinations from a grid of 4 caption styles x 5 BPMs x 5 keys (100 possible combos). Within each genre, lyrics were held constant while musical parameters varied, isolating the effect of BPM, key, and caption style on score.

Score Statistics

By Genre

Rank	Genre	Language	Mean	Best	Worst	Range
1	Punjabi/Bhangra	Punjabi	3.77	4.26	2.79	1.47
2	Electronic/EDM	Instrumental	3.71	5.29	2.80	2.49
3	Bollywood	Hindi	3.53	4.38	2.70	1.68
4	C-Pop	Chinese	3.20	4.47	2.13	2.34
5	Latin/Reggaeton	Spanish	3.19	3.90	2.36	1.54
6	Pop/Dance	English	3.05	4.31	1.98	2.33
7	Rock/Alternative	English	3.03	3.66	2.02	1.64
8	Hip Hop	English	2.92	3.38	2.52	0.86
9	Acoustic/Folk	English	2.63	3.31	2.03	1.28
10	R&B/Soul	English	2.62	3.21	2.14	1.07
--	Banger-optimized	Mixed	3.48	5.29	2.01	3.28

Top 10 Overall (out of 230 songs)

Rank	Score	Genre	BPM	Key	Caption Style
#1	5.29	EDM	130	Eb minor	Melodic techno, atmospheric, driving
#2	5.29	Banger Run	128	D minor	Dark electronic, industrial, driving
#3	4.90	EDM	138	F minor	Deep house, groovy bassline
#4	4.87	EDM	126	Eb minor	Progressive house, euphoric drop
#5	4.66	EDM	138	Bb minor	Melodic techno, dark and dreamy
#6	4.47	C-Pop	120	A minor	Emotional ballad, piano-driven
#7	4.40	Banger Run	128	Bb minor	Progressive house, festival anthem
#8	4.38	Bollywood	120	E minor	Club-ready beat, heavy bass
#9	4.36	Banger Run	130	D minor	Deep house, groovy bassline
#10	4.31	Pop/Dance	128	D major	Electronic pop, shimmering arpeggios

Best vs Worst Examples

The highest-scoring songs share common traits: driving four-on-the-floor beats, minor keys, 126-138 BPM, electronic/dark textures. The lowest-scoring songs tend to be slow R&B, acoustic folk, or jazzy hip hop with mellow dynamics.

Musical Analysis

Optimization Impact

The banger-optimized run (30 songs) used only parameter combinations that scored highest in the 200-song random sweep: dark electronic/industrial captions, 126-138 BPM, minor keys only.

Metric	Random (200 songs)	Optimized (30 songs)	Change
Mean score	3.17	3.48	+10%
Songs >= 3.5	20%	60%	3x
Songs >= 4.0	5%	20%	4x
Top score	5.29	5.29	Same

Data-driven parameter selection raised the floor and consistency without raising the ceiling. The winning formula: dark electronic/industrial styles, 128-130 BPM, D minor or Bb minor.

Key and Tonality

Minor keys consistently outperformed major keys across all genres. All top-5 overall songs were in minor keys.

BPM

Clear BPM sweet spots per genre. Slower BPMs (< 85) consistently underperformed. The highest-scoring region across all genres: 126-138 BPM.

Caption Style

The caption prompt significantly affects scores. Descriptive, production-oriented captions ("melodic techno, atmospheric pads, driving beat") outperformed vague ones. The "traditional/ethnic" keyword cluster (dhol, tumbi, bhangra) achieved the highest style-level mean (3.8).

Generated vs Training Data

AI-generated songs cluster tightly around the FMA training mean (3.17 generated vs 3.27 FMA), while real music has a long right tail up to 10. Even optimized generation peaks at 5.29 (67th percentile). The scorer correctly identifies that AI music does not yet match the quality distribution of human-made music.

Use Cases

• AI music quality research. Study what musical parameters (BPM, key, style, language) lead to higher/lower quality scores. The structured parameter grid makes controlled comparisons possible.
• Listening comparison. Download the top-5 and bottom-5 from each genre to hear what the scorer prefers. Do you agree with the model?
• Train better scorers. Use these scored songs as additional training data or evaluation benchmarks for music quality models.
• ACE-Step evaluation. Benchmark ACE-Step 1.5 across genres and languages with quantitative quality scores.
• Preference modeling. Use the ranked pairs (best vs worst within each genre) to train preference models or reward models for music generation.

Limitations

• Scorer bias. All scores come from a single model trained on FMA play counts. The model favors high-energy, beat-driven music and underrates mellow genres. Scores reflect learned statistical patterns, not absolute quality.
• Single generator. All songs are from ACE-Step 1.5. Results may not generalize to other music generation models.
• Limited seeds per combo. Each parameter combination was tested with only 1 random seed. Different seeds produce different songs from the same parameters.
• Lyrics held constant per genre. Within each genre test, all 20 songs share the same lyrics. Only musical parameters (BPM, key, caption style) vary.
• 2-minute songs. All songs are approximately 2 minutes. Longer or shorter durations may produce different quality characteristics.

How the Scores Were Computed

Each song was:

1. Loaded at 24kHz mono
2. Truncated to 30 seconds
3. Encoded through MERT-v1-330M to produce a 1024-dim embedding
4. Scored by the banger scorer MLP (1024 -> 512 -> 256 -> 128 -> 1)
5. Clamped to the 0-10 range

The scorer model (MAE 0.858, Spearman 0.468 on FMA test set) was trained on pre-computed MERT embeddings from treadon/fma-mert-embeddings.

Genre Test Details

Test	Genre	BPMs	Keys	Language	Lyrics Theme
01	Hip Hop	78, 85, 90, 95, 100	C/Bb/D/E/Ab minor	English	East coast, gritty, street
02	Pop/Dance	110, 118, 124, 128, 135	C/G/D/A/F major	English	Upbeat, summer, dance
03	R&B/Soul	65, 72, 80, 88, 95	Eb/Ab/Db maj, Bb/F min	English	Romantic, intimate
04	Latin/Reggaeton	90, 95, 100, 105, 110	A/D/E/G/C minor	Spanish	Party, dancing, heat
05	Bollywood	100, 110, 120, 130, 140	C/D/A/E min, G maj	Hindi	Film dance number
06	Punjabi/Bhangra	95, 105, 115, 125, 135	G/D/A/C maj, E min	Punjabi	Celebration, dance
07	C-Pop	75, 90, 105, 120, 130	C/G/F maj, A/D min	Chinese	Modern pop, emotional
08	Rock/Alt	120, 130, 140, 150, 160	E/A/D/G/B minor	English	Indie, garage, alt
09	Electronic/EDM	122, 126, 130, 138, 150	A/F/C/Eb/Bb minor	Instrumental	House, techno, trance
10	Acoustic/Folk	85, 95, 105, 115, 125	G/C/D/A/E major	English	Singer-songwriter
--	Banger Run	126, 128, 130, 135, 138	Eb/F/Bb/C/D minor	Mixed	Optimized for score

Citation

@misc{acestep2025,
  title={ACE-Step: A Step Towards Music Generation Foundation Model},
  author={ACE-Step Team},
  year={2025},
  url={https://huggingface.co/ACE-Step/Ace-Step1.5}
}

@article{li2023mert,
  title={MERT: Acoustic Music Understanding Model with Large-Scale Self-supervised Training},
  author={Li, Yizhi and Yuan, Ruibin and Zhang, Ge and Ma, Yinghao and others},
  journal={arXiv preprint arXiv:2306.00107},
  year={2023}
}

Dataset Card Contact

treadon on HuggingFace