Symbolic Music Generation
Symbolic generation models notes and events, producing editable MIDI or notation.
Related
Symbolic generation never touches a waveform. It treats pitch, duration and velocity as words and learns to write the next string.
The problem
Audio generation cannot easily change one note; symbolic generation can. Yet the voices a model writes lack direction and sound like studies.
How it works
MIDI becomes tokens, with pitch, duration and velocity encoded separately, and a Transformer or diffusion model learns context. Conditions can be style, chord progression, motif or section markers. Length is limited by the context window, so long pieces need continuation and careful seams.
Practice and judgement
Supply a chord skeleton or motif, then let the model continue. Limit range and polyphony, generate several takes and combine. Edit velocity, legato and orchestration in the DAW, and keep seeds and settings for reproduction. Judge a take by cadences, a bass that does more than repeat roots, and inner voices with their own motion.
What to listen for
Hear where phrases breathe, whether cadences resolve and whether the motif changes on return. Then hear the bass; if it only stamps roots, the harmony cannot stand.
Common misunderstandings
Symbolic generation is composition; more data is always better; MIDI output needs no editing. The model supplies material; phrasing and orchestration remain human decisions.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Sources
- 符号音乐表示与 token 化方法的公开研究，依 ISMIR 论文核查，未抄文本
- 音乐生成评估中和声、动机与节奏的常用判据，依公开论文核查，未抄文本
Used for fact-checking only; all prose is original.