MuseCoco
Microsoft's text-to-MIDI composition copilot: a two-stage model that turns text into editable symbolic music via musical attributes.
About
MuseCoco (Music Composition Copilot) tackles text-to-music in the symbolic domain, where the output is editable MIDI rather than a frozen waveform. It splits the problem in two: a BERT-large model turns the text description into explicit musical attributes (key, tempo, emotion, bar count, time signature), and a Linear Transformer then generates the music conditioned on those attributes, trained self-supervised on roughly 947k MIDI files. That design gives precise control over named attributes instead of vibes.
The authors report it beats baselines by at least 1.27 on musicality, 1.08 on controllability, and 1.32 overall, with about a 20% lift in objective control accuracy, and they trained a 1.2B-parameter large-scale variant. Code and checkpoints live in Microsoft's MIT-licensed muzic repository, making it one of the most capable open symbolic-generation models in the catalog. The paper is from June 2023.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (arXiv 2306.00110): >=1.27 musicality / 1.08 controllability / 1.32 overall score gains over baselines; ~20% improvement in objective control accuracy
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: paper (arXiv 2306.00110, June 2023) from Microsoft Research authors with full experiment tables; code and checkpoints in github.com/microsoft/muzic, confirmed MIT-licensed via multiple third-party attributions. Strengths: strong reported controllability gains, self-supervised data-efficient design, permissive license, reputable authors. Docked for: (1) evaluation is author-run on a synthesized 5k-pair test set, not independent; (2) large 1.2B variant is compute-heavy for self-hosting. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Paper arXiv 2306.00110 (June 2023), Microsoft Research
- · 1.2B-param large variant; trained on ~947k MIDI files; >=1.27 musicality / 1.08 controllability / 1.32 overall gains over baselines (author-reported)
- · Code + checkpoints in microsoft/muzic repo