0MGE: Neural Granular Engine

AI music generation from YOUR music. Scan, train, generate — all locally. No cloud, no API, no subscription.

Pre-trained neural granular engine trained on 2389 tracks (~48 hours of music). Generates new drone landscapes, textures, and atmospheres from a learned grain field.

Full source code, training pipeline, and desktop app: 0MGE on GitHub

Listen to demo — audio-reactive player with GLSL shader

---

What This Is

A neural network that learns from music and generates new sound. Not text-to-music. It scans audio files, cuts them into millions of micro-grains, and trains a navigator to assemble those grains into new sonic worlds.

This repo contains a pre-trained model and grain pool trained on Slut Online's music (2389 tracks). Download, generate, hear what it sounds like.

To generate from your own music — use the desktop app (scans your library, trains locally).

---

Generate

git clone https://github.com/0penAGI/0MGE.git && cd 0MGE
pip install numpy torch librosa scikit-learn soundfile

python3 granular_field.py --pool granular_pool_v2_int16.npz --model granular_multi_v1.pt --bars 60 --multi-stream --seed 42

Output: granular_output/granular_60bars_*.wav (stereo, 22050 Hz).

---

Files

File	Size	Description
granular_multi_v1.pt	6.4 MB	FP32 6-stream navigator with attractor field
granular_multi_v1_int8.npz	1.6 MB	Navigator INT8 quantized (weights only, not the grain pool)
granular_multi_v1_int8_meta.json	0.4 KB	INT8 scale metadata
granular_pool_v2_int16.npz	5.8 GB	Full grain pool, 566K grains, INT16 with per-row peak normalization
granular_pool_lite.npz	64 MB	Features only (22-dim), no raw audio

---

Architecture

MultiNavigator — Transformer (4 heads, 3 layers, 192 hidden). 48-dim state, 12-step context. 6 independent stream heads select grains via softmax over pool similarities. Attractor field (z0-inspired) provides learned global state per stream for long-range coherence — pulls generation toward meaningful musical directions instead of random walk.

Stream	Band	Role
sub	20–120 Hz	Low-frequency energy
drums	120–500 Hz	Transient body
harmonic	500–2000 Hz	Tonal content
texture	2–4 kHz	Mid-high presence
noise	4–8 kHz	High-frequency detail
air	8–11 kHz	Upper spectrum

---

Grain Pool

Three-tier hierarchy extracted via STFT (n_fft=1024, hop=256):

Level	Duration	Count
Micro (μ)	~55ms	425K
Meso (σ)	~300ms	118K
Macro (Ω)	~3s	23K
Total	—	566K

22-dimensional spectral features per grain. MiniBatchKMeans clustering (1024 clusters).

---

Quantization

Metric	FP32	INT8	Delta
Critic score	0.292	0.285	0.008
File size	5.3 MB	1.4 MB	3.9×

---

Audio Samples

Sample	Duration	Seed	Critic
samples/drone-01.mp3	60s	42	0.375
samples/drone-02.mp3	60s	1337	0.371
samples/drone-03.mp3	60s	2026	0.400
samples/drone-04.mp3	60s	777	0.374
samples/drone-05.mp3	60s	314	0.355
samples/drone-06.mp3	60s	256	0.376

---

Citation

@software{0mge2026,
  title={0MGE: Neural Granular Engine},
  author={0penAGI},
  year={2026},
  url={https://github.com/0penAGI/0MGE}
}

License

MIT

Trained on music by Slut Online with permission.