Symbolic MIDI Generation

It writes notes rather than audio, editable and transposable, but the structure is yours to finish.

Model profile

Vendor

OpenAI · Google Magenta

Form

本地模型

Input

MIDI · 风格提示 · 乐器组合

Output

MIDI

Licence

各项目的代码与权重许可不同，训练 MIDI 的来源与商用范围需按仓库说明核对

Verified

2026-10-30

Related

MIDI ProtocolPhrasePiano

What it lets you finish

Continue four to eight bars and hand you an opening worth editing; add harmony, a bass line or drums under a melody; translate a passage into another instrument group; generate dozens of motifs to choose from; produce style examples for teaching; export multi-track MIDI and assign sounds yourself in the DAW.

Inputs and outputs

You input a MIDI fragment plus style and instrument tags, and some models accept sampling parameters such as temperature; the output is a MIDI file or a live note stream, usually multi-track and quantised. No audio comes out: sounds, velocities and articulation are added in the DAW or a sample library. Length is counted in bars, so longer pieces are continued in sections and joined.

Strengths

• The result is notes, so changing one costs nothing and editing stays cheap.
• Small files fit version control and open directly in notation software.
• It writes several parts at once, which lets you judge voice leading rather than timbre alone.
• Good for motifs, bass lines and drum drafts, faster to spread out than writing by hand.
• Low training and inference cost, and local runs need no network.
• Fits DAWs and notation tools; quantised fragments drop straight into a project.

Weaknesses and failure modes

• Long form: after a few dozen seconds it repeats or wanders, and transitions between sections are yours to write.
• Harmony: parallel fifths, voice crossing and out-of-range notes appear, especially in four-part writing.
• Rhythm: output is snapped to the grid, so the small push and pull of syncopation and swing is flattened and drums sound stiff.
• Expression: velocity, legato and pedalling are mostly missing, and the same notes read like typing.
• Style: it learns statistical resemblance, not the rules of writing, and imitation shows at the details.
• Rework: the share of usable bars is low, and the time spent auditioning and trimming is routinely underestimated.

Terms, price and alternatives

Questions to check: are the weights public, and under what licence? Does the training MIDI include copyrighted works? How is copyright in generated MIDI assigned? Is credit required for commercial use? Are there limits on passing generated fragments to clients?

Price: open projects are free with your own compute; hosted services bill per subscription or per use, as stated on the official pages.

Alternatives: AIVA when you want finished audio and orchestration without writing parts; a harmony tool such as Scaler when you want a controlled chord progression rather than generation; a demo such as MuseNet when you only want to hear where a style goes.

Sources

• 符号音乐生成的代表项目与输出形态，依公开仓库与论文摘要核对，未抄原文
• MIDI 生成结果的后期修剪，依编曲实践通识核对，未抄文本

Used for fact-checking only; all prose is original.