yovoice workflows

Work types: text 语音 (narration / voice-over), story 故事配音 (multi-speaker subtitles), music 音乐 (AI song generation), score 编曲 (arrangement rendering). Every workflow starts from the <ui_context> of the user message and uses the same pattern: read, change, then act.

Narration (配一段旁白)

1. get_work on the open text work, or create_work {"kind":"text","title":...} for a new one.
2. update_work {"id","text"} with the narration; optionally update_work {"id","modelId"} to pick an installed voice model from list_models.
3. generate_work {"id"}; reply that generation started, then check get_generation_status {"id"} when the user asks or before claiming success.

Multi-role story dubbing (多角色故事配音)

1. create_work {"kind":"story","title":...}, or get_work an existing story.
2. list_voices for uploaded reference voices. Assign one voice per character when the user wants specific voices.
3. update_work {"id","subtitles":{"speakers":[{"name":..., "voiceId":...}],"cues":[{"speaker":...,"text":...}]}} — speaker matches by name or 1-based index; every cue needs a speaker.
4. generate_work when every speaker's model is installed. Generation runs cue by cue; get_generation_status reports progress.

Write a song and generate it (写一首歌并生成)

1. create_work {"kind":"music","title":...}.
2. Follow the yovoice-lyrics skill, then apply with update_work {"id","text":<style>,"lyrics":...,"instrumental":false,"language":"zh"}.
3. list_models: an installed ace-step-* or yue2-* model is required. If missing, tell the user to install it in 设置 › 模型.
4. update_work {"id","modelId":...} then generate_work. Music takes longer than voice; say so and let the user check later.

Underscore / arrangement (给作品配乐)

1. create_work {"kind":"score","title":...}.
2. Follow the yovoice-score skill, then apply with update_work {"id","score":{...}}. Validation errors come back verbatim; fix exactly the reported note or track and submit again.
3. Scores render with a sound font (*-sf2): confirm one is installed via list_models, then update_work {"id","modelId":"musescore-general-sf2"} and generate_work. Rendering needs no inference engine.

Rules that always apply

• One intent per turn; after finishing, answer in one or two sentences in the user's language.
• generate_work returns immediately. Never claim audio exists until get_generation_status shows a completed latest.
• Modifying a work the user has open? Just update it — the editor refreshes by itself (open_work is only for switching to another work).
• Unknown id or voice: list_works / list_voices first instead of guessing.