🎵 Amadeus: Autoregressive Model with Bidirectional Attribute Modelling for Symbolic Music

Amadeus is a novel symbolic music (MIDI) generation framework. We use autoregressive modeling for note sequences, discrete diffusion models for intra-note attributes, and representation optimization to enhance model performance. Compared to current mainstream autoregressive or hierarchical autoregressive models, Amadeus achieves significant improvements in generation quality, speed, and controllability. While significantly improving generation quality, we have achieved a speedup of at least 4x compared to pure autoregressive models. We also support a training-free fine-grained attribute control mechanism, which endows Amadeus with maximum flexibility. We will continuously update the code, models, and datasets.

---

🏗️ Model Architecture

---

📅 Changelog

• 2025-08-28: Released inference code and the Amadeus-S model

---

⚙️ Installation and Usage

Set up the environment (inference only):

!!! 训练使用environment yml创建环境 conda env create -f environment.yml

conda create -n amadeus_slim python=3.10

conda activate amadeus_slim

pip install -r demo/requirements.txt

First run:

# Chinese interface

python demo/Amadeus_app_CN.py

# English interface

python demo/Amadeus_app_EN.py

Note:

Amadeus_app_CN.py

is for the Chinese interface, and

Amadeus_app_EN.py

is for the English interface.

👉 The model will be automatically downloaded to the models/ folder, which includes a usable soundfont. Please modify the path of DEFAULT_SOUND_FONT in Amadeus/symbolic_encoding/``midi2audio.py.

Example of command-line generation:

python generate.py -wandb\_exp\_dir models/Amadeus-S -text\_encoder\_model google/flan-t5-base -temperature 2 -prompt "A lively and melodic pop rock song featuring piano, overdriven guitar, electric drum and electric bass, set in a fast 4/4 tempo and the key of C# minor, with a frequently recurring chord progression of D, A, C#m, and F# that evokes a mix of emotion and love."

---

📂 Repository Structure

Amadeus/

├── demo/                   # Example scripts and interfaces (CN/EN)

├── Amadeus/                # Core model and symbolic encoding

├── assets/                 # Architecture diagrams and sample audio files

├── data\_representation     # Data processing

├── models/                 # Downloaded or cached pre-trained models

└── generate.py             # Command-line generation entry point

---

📊 Evaluation Results

We evaluated generation speed, text alignment, and note attribute control accuracy on the MidiCaps dataset. The results are as follows:

Model	Speed (notes/s)	CLAP ↑	TBT ↑	CK ↑	CTS ↑	CI ↑	CMtop3 ↑
Text2Midi	4.02	0.19	31.76	22.22	84.15	19.92	60.57
MuseCoco	1.67	0.19	34.21	14.66	94.24	22.42	38.18
T2M-inferalign	4.02	0.20	39.32	29.80	84.32	20.13	47.74
Amadeus	16.23	0.20	73.93	39.31	96.98	26.01	65.52
Amadeus-M	10.51	0.21	76.31	43.07	97.02	27.11	66.39

---

🤝 Acknowledgements and Contributions

The development of Amadeus is inspired by the music and AI communities, with the goal of serving music creators, not replacing them.

We welcome developers and researchers to contribute code or provide suggestions — please reach out to us via Issues or Pull Requests.

Part of the design of this project references JudeJiwoo/nmt, and we would like to express our gratitude here 🙏.

---

⚠️ Notes

The current model is relatively small and may not always generate MIDI that fully matches the description.
You can try slightly adjusting parameters such as temperature or top-p to improve the results.

We will continue to improve the model to provide more stable and higher-quality generation.

---

📚 Citation

If you find Amadeus helpful for your research or create，please cite our paper:

@article{su2025amadeus,
  title   = {Amadeus: Autoregressive Model with Bidirectional Attribute Modelling for Symbolic Music},
  author  = {Su, Hongju and Li, Ke and Yang, Lan and Zhang, Honggang and Song, Yi-Zhe},
  journal = {arXiv preprint arXiv:2508.20665},
  year    = {2025}
}