YuE2 music (YuE2 音乐)
YuE2 renders a song from a style prompt, lyrics and, optionally, a two-voice melody score (abc field: Vocal with chords, Ins instrumental melody). Call get_work first: text is the style, settings.lyrics the lyrics, settings.abc the score, settings.instrumental the instrumental switch, settings.modelOptions.yue2 the generation settings.
Style and lyrics
- Style in English, up to 2000 characters: sung language, genre, lead vocal, core instruments and tempo, plus a sentence or two on how the song unfolds (how it opens and when the voice enters, what changes between sections, how it ends). Write style and lyrics with the yovoice-lyrics skill; its song-forms reference has the YuE2 style guidance, and its yue2-examples reference distils the official demos, the paper's findings and the official editing walkthrough. Put tempo and meter in the score too and keep them consistent.
- Lyrics: section tags on their own lines (
[Verse], [Chorus], [Bridge], [Outro], optionally with a short note such as [Verse 2 – whispered]) followed by the sung words. Only words to be sung; no notes or instructions. Song length follows the lyrics.
- Each
plan_melody call writes a new score with a fresh seed. When the user wants a different tune, opening or structure, plan again (or change the style's shape) rather than generating more takes from the same score: takes from one score share its melody and form.
Pick the workflow
| Request |
What to do |
A song in score mode (the default: cot is not off, or instrumental) |
update_work style + lyrics → plan_melody → get_generation_status until melody appears → update_work {"abc": melody.abc}. Then stop and ask the user to review: the score is in the 旋律谱 section, where they can read it and press 试听 (listen). Call generate_work only after they approve or ask for changes you have made. generate_work refuses a score-mode YuE2 work without a score. |
| The user explicitly wants to skip reviewing the score (直接生成) |
generate_work {"id","withoutScore":true}: YuE2 writes the score itself while singing, and it is filled into the work afterwards. |
Free mode (cot is off) |
update_work style + lyrics, then generate_work; there is no score to review. |
| Cover a song (翻唱) |
cover_melody {"id","path"} with a local audio file (needs SheetSage2; propose installing it otherwise) → apply its melody.abc → new style and lyrics → generate_work. Keeps the melody, not the singer's voice. |
| Instrumental (纯音乐) |
update_work {"instrumental": true}; lyrics optional (section tags only). Write and review the score first as in the first row; generating moves the vocal melody to Ins and plays it. transform_melody {"op":"instrumental"} converts the stored score in place. |
| Skip the score |
update_work settings {"modelOptions":{"yue2":{"cot":"off"}}}; the take then has no editable score. "auto" goes back to score-based generation. |
| Keep the melody, let YuE2 re-arrange |
transform_melody {"op":"stripChords"}. Change the tempo with {"op":"tempo","tempo":96}. |
| Change harmony, style, lyrics or structure |
Read references/editing.md first. |
| Write or fix notation |
Read references/abc-dialect.md before writing any ABC. |
Reference files are in this skill's folder (from the workspace: .agents/skills/yovoice-music/references/); read only the one the task needs.
Settings
modelOptions.yue2: cot (auto | full | melody | off; auto picks full with chords, melody without), candidates 1-4 (1; how many takes one generate makes, each with the next seed), num_inference_steps 1-64 (8; the official baseline is 32, which can sound cleaner but takes longer), abc_temperature 0.05-2 (0.7, score writing), abc_top_p 0.01-1 (0.9), abc_top_k 1-500 (30), abc_repetition_penalty 0.5-3 (1.005; keep it near 1, a score repeats bars and motifs by design), semantic_temperature 0.05-2 (1.0), semantic_top_p 0.01-1 (0.95), semantic_top_k 1-500 (100), semantic_repetition_penalty 0.5-3 (1.2), ar_lora / nar_lora (an imported adapter ID or empty; set in the app, not by you), ar_lora_scale / nar_lora_scale -10..10 (1). Change them only when asked; the defaults are the tuned route.
Stable Audio sound effects
For an sfx work, write text as a concrete English description of the sound source and its action, not a music style. Include useful spatial distance or indoor/outdoor placement and the materials involved, for example: a heavy wooden door closing in a small tiled hallway, close microphone, short natural reverberation or rain hitting a metal roof outdoors, distant traffic far behind. Request 1–30 seconds. Avoid genre, band, melody, tempo, singing and other music-style wording. Small SFX is best for short effects; Medium trades a larger model for a better balance of duration, detail and quality.
Be honest about results
YuE2 follows a score loosely: it does not guarantee every note, exact timing, or that an instrument named in the style is absent. A valid score or a finished take is not proof it sounds right; say what you changed and invite the user to listen. Every edit becomes a new take; earlier takes stay in the history for comparison.
1---2name: yovoice-music3description: YuE2 songs and instrumentals in yovoice — melody scores (write, extract from a song, edit, convert to instrumental), generation mode, and musical edits such as reharmonizing, changing style or lyrics while keeping the melody. Use whenever a music work uses a yue2- model.4metadata:5 source: https://skillmd.com/skills/leemysw/yovoice-music6---78# YuE2 music (YuE2 音乐)910YuE2 renders a song from a style prompt, lyrics and, optionally, a two-voice melody score (`abc` field: `Vocal` with chords, `Ins` instrumental melody). Call `get_work` first: `text` is the style, `settings.lyrics` the lyrics, `settings.abc` the score, `settings.instrumental` the instrumental switch, `settings.modelOptions.yue2` the generation settings.1112## Style and lyrics1314- Style in English, up to 2000 characters: sung language, genre, lead vocal, core instruments and tempo, plus a sentence or two on how the song unfolds (how it opens and when the voice enters, what changes between sections, how it ends). Write style and lyrics with the yovoice-lyrics skill; its song-forms reference has the YuE2 style guidance, and its yue2-examples reference distils the official demos, the paper's findings and the official editing walkthrough. Put tempo and meter in the score too and keep them consistent.15- Lyrics: section tags on their own lines (`[Verse]`, `[Chorus]`, `[Bridge]`, `[Outro]`, optionally with a short note such as `[Verse 2 – whispered]`) followed by the sung words. Only words to be sung; no notes or instructions. Song length follows the lyrics.16- Each `plan_melody` call writes a new score with a fresh seed. When the user wants a different tune, opening or structure, plan again (or change the style's shape) rather than generating more takes from the same score: takes from one score share its melody and form.1718## Pick the workflow1920| Request | What to do |21| --- | --- |22| A song in score mode (the default: `cot` is not `off`, or instrumental) | `update_work` style + lyrics → `plan_melody` → `get_generation_status` until `melody` appears → `update_work` {"abc": melody.abc}. Then stop and ask the user to review: the score is in the 旋律谱 section, where they can read it and press 试听 (listen). Call `generate_work` only after they approve or ask for changes you have made. `generate_work` refuses a score-mode YuE2 work without a score. |23| The user explicitly wants to skip reviewing the score (直接生成) | `generate_work` {"id","withoutScore":true}: YuE2 writes the score itself while singing, and it is filled into the work afterwards. |24| Free mode (`cot` is `off`) | `update_work` style + lyrics, then `generate_work`; there is no score to review. |25| Cover a song (翻唱) | `cover_melody` {"id","path"} with a local audio file (needs SheetSage2; propose installing it otherwise) → apply its `melody.abc` → new style and lyrics → `generate_work`. Keeps the melody, not the singer's voice. |26| Instrumental (纯音乐) | `update_work` {"instrumental": true}; lyrics optional (section tags only). Write and review the score first as in the first row; generating moves the vocal melody to Ins and plays it. `transform_melody` {"op":"instrumental"} converts the stored score in place. |27| Skip the score | `update_work` settings `{"modelOptions":{"yue2":{"cot":"off"}}}`; the take then has no editable score. `"auto"` goes back to score-based generation. |28| Keep the melody, let YuE2 re-arrange | `transform_melody` {"op":"stripChords"}. Change the tempo with {"op":"tempo","tempo":96}. |29| Change harmony, style, lyrics or structure | Read [references/editing.md](references/editing.md) first. |30| Write or fix notation | Read [references/abc-dialect.md](references/abc-dialect.md) before writing any ABC. |3132Reference files are in this skill's folder (from the workspace: `.agents/skills/yovoice-music/references/`); read only the one the task needs.3334## Settings3536`modelOptions.yue2`: `cot` (`auto` | `full` | `melody` | `off`; auto picks full with chords, melody without), `candidates` 1-4 (1; how many takes one generate makes, each with the next seed), `num_inference_steps` 1-64 (8; the official baseline is 32, which can sound cleaner but takes longer), `abc_temperature` 0.05-2 (0.7, score writing), `abc_top_p` 0.01-1 (0.9), `abc_top_k` 1-500 (30), `abc_repetition_penalty` 0.5-3 (1.005; keep it near 1, a score repeats bars and motifs by design), `semantic_temperature` 0.05-2 (1.0), `semantic_top_p` 0.01-1 (0.95), `semantic_top_k` 1-500 (100), `semantic_repetition_penalty` 0.5-3 (1.2), `ar_lora` / `nar_lora` (an imported adapter ID or empty; set in the app, not by you), `ar_lora_scale` / `nar_lora_scale` -10..10 (1). Change them only when asked; the defaults are the tuned route.3738## Stable Audio sound effects3940For an `sfx` work, write `text` as a concrete English description of the sound source and its action, not a music style. Include useful spatial distance or indoor/outdoor placement and the materials involved, for example: `a heavy wooden door closing in a small tiled hallway, close microphone, short natural reverberation` or `rain hitting a metal roof outdoors, distant traffic far behind`. Request 1–30 seconds. Avoid genre, band, melody, tempo, singing and other music-style wording. Small SFX is best for short effects; Medium trades a larger model for a better balance of duration, detail and quality.4142## Be honest about results4344YuE2 follows a score loosely: it does not guarantee every note, exact timing, or that an instrument named in the style is absent. A valid score or a finished take is not proof it sounds right; say what you changed and invite the user to listen. Every edit becomes a new take; earlier takes stay in the history for comparison.