arjunbahuguna/moonbeam-microtonal

Moonbeam Microtonal (1-cent resolution, 1200 pitch classes/octave)

Domain adaptation of Moonbeam MIDI Foundation Model for microtonal music with uniform cent-resolution pitch grid.

Model Details

Property	Value
Base Model	Moonbeam S (309M params)
Tuning System	1200-TET (1200 pitch classes/octave)
Cents per Step	1.00
Total Parameters	~312M
Architecture	LLaMA + MRA + GRU sub-decoder
GRU Vocab Size	3529

Key Modifications from Base Moonbeam

1. Pitch class vocabulary expanded from 12 → 1200 (uniform 1-cent grid)
2. GRU output head expanded with interpolated initialization from pretrained weights
3. FME (Fundamental Music Embedding) handles microtonal pitch natively via continuous sinusoidal embedding
4. MRA (Multidimensional Relative Attention) encodes relative pitch/onset/duration

Microtonal Traditions in Training Data

• Arabic Maqam (quarter-tones: Rast, Bayati, Hijaz, Saba, Sikah, Nahawand)
• Turkish Makam (53-comma: Rast, Ussak, Huseyni, Hicaz, Segah)
• Indian Raga (22-Shruti: Bilawal, Kafi, Bhairavi, Yaman, Todi, Marwa)
• Indonesian Gamelan (Slendro, Pelog)
• Western microtonal (just intonation, meantone, quarter-tone chromatic)
• Standard 12-TET (backward compatibility)

Architecture

Each musical event = compound token: (onset, duration, octave, pitch_class, instrument, velocity)

• pitch_class ∈ {0, 1, ..., 1199} — each value = 1 cent
• octave ∈ {0, 1, ..., 10} — 11 octaves
• FME input accepts continuous float values (handles sub-cent precision)
• Total pitch resolution: 13200 unique pitches (1200×11 octaves)

Citation

@article{guo2025moonbeam,
  title={Moonbeam: A MIDI Foundation Model},
  author={Guo, Zixun and Dixon, Simon},
  journal={arXiv preprint arXiv:2505.15559},
  year={2025}
}