Magenta
A research toolkit for MIDI and audio, built for reproducible experiments.
Model profile
- Vendor
- Google Magenta
- Form
- 开源工具集
- Input
- MIDI · 音频 · 文本
- Output
- MIDI · 音频
- Licence
- 代码以 Apache 2.0 为主，各模型与预训练权重分别标注，商用前看具体仓库
- Verified
- 2026-10-31
Related
A toolbox, not a song
Magenta is a Google research project, not a one-line-to-song product. It provides a set of models and tools: MusicVAE for melody interpolation and recombination, Melody RNN for continuation, Drums RNN for drum parts, Performance RNN for performance dynamics, DDSP for differentiable audio synthesis and NSynth for timbre. Each piece is small; together they let you assemble your own generation pipeline.
MIDI in, MIDI out
Most tools take MIDI and return MIDI: pitch, duration and velocity stay explicit, so editing feels like editing a score rather than audio. DDSP and NSynth are the exceptions, returning audio with adjustable timbre but no words. Magenta does not handle lyrics or singing; that needs another model. The division of labour makes it handy for arranging, motif development and drum variation.
| Tool | Input | Output | Good for |
|---|---|---|---|
| MusicVAE | MIDI fragment | MIDI | Interpolation, continuation |
| Melody RNN | Melody | MIDI | Continuation |
| Drums RNN | Style or fragment | MIDI drums | Variation |
| DDSP | Audio or MIDI | Audio | Timbre transfer |
The value is reproducibility
The same MIDI, checkpoint and random seed can be run again. For research or teaching that matters more than "sounds good": change one variable and compare melodic direction, drum density or timbral brightness. Magenta Studio wraps these models as Ableton plugins for people who do not code; for batch work, use the Python package. Output drops into a DAW for further arranging or as a sound source.
Maintenance risk after archiving
The main repository is now read-only and some tools have moved to separate repositories, so install notes and dependency versions often disagree. A changed TensorFlow version, checkpoint or Python dependency can stop the whole thing. There is no support desk and no stable API; to put it into production, freeze an environment you can keep. New work still appears, but not at the pace of a maintained product.
Licence, maintenance and alternatives
Code is mostly Apache 2.0; models and pretrained weights are labelled separately, so read the repository before commercial use. Alternatives: Suno or YuE for finished songs; Stable Audio Open for textures and effects; newer symbolic projects such as ACE-Step if you need active maintenance. Magenta's advantage is transparency and modifiability, not audio quality.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Along the same line
Sources
- Google Magenta 项目页与 Magenta Studio 文档（工具清单、输入输出与维护状态）。
- MusicVAE、Melody RNN 与 DDSP 的公开论文和代码说明。
- 主仓库归档与独立仓库开发的公开记录。