leemysw/yovoice/yovoice-product

yovoice product guide

yovoice runs every model locally through its inference engine. Models and the engine are downloaded in Settings; generation happens on this machine and works offline afterwards.

Pages (open_page keys)

The assistant opens with ⌘J / Ctrl+J, the launcher in the corner, or the title bar button.

Choosing a voice model

Hardware rule of thumb: Kokoro runs on any CPU; cloning models want 4–6 GB VRAM or 8 GB+ Apple Silicon; music models want 12 GB+ VRAM or 16 GB+ Apple Silicon. Q8 is the smallest precision with near-original quality; BF16/F16/ORIG are larger and slightly better.

Recommend in this order: what the user already has installed (list_models), then the lightest model that does the job.

Music and arrangement models

• ACE-Step 1.5 Turbo / XL Turbo: songs or instrumentals from style (≤ 512 chars) + lyrics (≤ 4000 chars, [verse] [chorus] tags). Duration auto or 10–300 s. XL is higher quality and much larger.
• YuE2 3B: vocal songs (length follows the lyrics) or instrumentals (writes a score, then instruments play the vocal melody; lyrics optional), non-commercial licence. Can plan an ABC melody score first or cover a song via SheetSage2 (yovoice-music skill).
• Stable Audio 3 Small Music / Small SFX / Medium: instrumental music (Small Music, Medium) or sound effects (Small SFX) from an English description; no lyrics or vocals. Duration 1–120 s (Small) or up to 380 s (Medium), default 30 s. The 生成方式 setting can also start from an existing take of the same work: 改编 Rework (edit = init_audio, init_noise_level 0–1, lower stays closer) or 局部重绘 Inpaint (edit = inpaint_audio, inpaint_start/inpaint_end seconds), with sourceGenerationId naming the take. Stability AI Community License (free under USD 1M yearly revenue).
• SheetSage2: extracts melody and chords from a song for YuE2 covers.
• Lyric Aligner (optional): on the playback page, Auto-align separates the vocals and aligns the lyrics word by word to a generated song; saved with the version.
• SoundFonts (MuseScore General, GeneralUser GS, SGM-V2.01) render arrangements on the score page.
• All music models need inference engine v0.9.1 or later.

Common fixes

• Dragging, repeating or trailing silence on long text (Breeze, VoxCPM2): shorten text_chunk_size in Advanced, then lower temperature to 0.6–0.8.
• Voice changes between chunks without a reference: design the voice on a short line, save it to the voice library, then clone from it.
• Cloning sounds wrong: use one clear speaker, little noise, 5–15 s; transcripts must match word for word for OmniVoice and Breeze.
• Audio check warnings (shown on a take, a story line or in inspect_audio): clipping / clicks / cutoff / silence / silent → regenerate that take or line; quiet → pick a target loudness when exporting (Audiobook −18, Podcast −16, Video −14 LUFS), no need to regenerate; dc → ignore.
• Slow or out of memory: pick Q8 or a smaller model, close other GPU apps, or switch compute device in Settings › Engine.
• Model missing or engine too old: the user installs it in Settings › Voice models, Settings › Music models (ACE-Step, YuE2, Stable Audio 3, SheetSage2) or Settings › Engine; you can propose it with propose_action.

Error codes

Errors reach you as stable codes (in <ui_context>, get_generation_status, get_recent_errors). The list below is the exact text the user sees in Chinese; explain it in the user's language and give the fix.