Create story dubbing, speech, full songs, and multi-track scores on your own computer. Runs locally, with your data under your control.
他脸上黑而且瘦，已经不成样子；穿一件破夹袄，盘着两腿，下面垫一个蒲包，用草绳在肩上挂住；见了我，又说道，“温一碗酒。”掌柜也伸出头去，一面说，“孔乙己么？你还欠十九个钱呢！”孔乙己很颓唐的仰面答道，“这……下回还清罢。这一回是现钱，酒要好。”掌柜仍然同平常一样，笑着对他说，“孔乙己，你又偷了东西了！”但他这回却不十分分辩，单说了一句“不要取笑！”“取笑？要是不偷，怎么会打断腿？”孔乙己低声说道，“跌断，跌，跌……”他的眼色，很像恳求掌柜，不要再提。此时已经聚集了几个人，便和掌柜都笑了。我温了酒，端出去，放在门槛上。他从破衣袋里摸出四文大钱，放在我手里，见他满手是泥，原来他便用这手走来的。不一会，他喝完酒，便又在旁人的说笑声中，坐着用这手慢慢走去了。
Swipe to explore the liner notes →
DEMO Listen freely
Story dubbing · Speech generation · Music generation · Score
Text, subtitles, lyrics, or notes for each part
Assign characters and choose models or General MIDI sound fonts
Keep every take as a version, compare them, and export
A few seconds of audio. A familiar voice.
Upload or record a reference clip to clone a voice, or describe a sound in words to design a brand-new tone from scratch.
你终于来了，我等你好久了。
Bright and excited, or quiet and restrained? Shape the delivery with emotion tags, reference audio, or natural-language prompts so the voice says more than words alone.
Available controls vary by modelMUSIC / 03Music generation demo with sound, click to play
Generate full songs with lyrics and style, or create instrumentals and sound effects with Stable Audio 3, with score-based arrangement, inpainting, and word-by-word lyric sync.
Music generation requires a supported inference engine; some models need a GPU or Apple SiliconSCORE / 04Score demo with sound, click to play
Write parts bar by bar with an intuitive grid editor, preview them in real time, and import or export standard MIDI files. Renders without an inference engine, wherever you are.
Score rendering uses General MIDI sound fonts and does not rely on a GPU
No waiting in line, and no privacy concerns. All inference happens locally, your assets stay on disk, and you can create even when offline.
An intuitive desktop app out of the box, and a scriptable engine underneath. Use the CLI, HTTP API, or MCP to integrate yovoice directly into AI Agent workflows.
| Model | Main use cases and core capabilities |
|---|---|
| IndexTTS 2.0 / 2.5 | Voice cloning and emotional control: 2.0 covers Chinese and English, while 2.5 adds multilingual support, reference performance, and pronunciation editing |
| VoxCPM2 | Voice design from text descriptions, reference-based cloning, and fine-grained cloning assisted by reference transcripts |
| OmniVoice | Attribute-based voice design, voice cloning, and non-verbal speech tags such as laughter and sighs |
| Qwen3-TTS | Multilingual voice cloning, 9 built-in voices with style direction, and voice design from natural-language descriptions |
| Breeze TTS 2 | Chinese and English voice design, cloning, and expressive delivery; GPU recommended |
| Kokoro-82M 1.0 / 1.1-zh | Lightweight preset voices: 49 in 1.0 Official, 100 Chinese voices in experimental 1.1-zh; supports voice blending |
| ACE-Step 1.5 Turbo / XL Turbo | Turbo creates songs or instrumentals from styles and lyrics, supporting 50+ vocal languages; XL Turbo offers higher quality |
| YuE2 3B | Generates full vocal songs from lyrics and styles with lyric-defined lengths; supports editable melody scores |
| Stable Audio 3 | Generate instrumental music (up to 6 minutes) and short sound effects (1–30 s) from English descriptions, with rework and inpainting; Stability AI Community License |
| Ming 3.0 / XTTS v2 / Voxtral / Drama 3 | Ming 3.0 covers 7-language voice design and cloning; XTTS v2 offers 17-language cloning; Voxtral provides 20 built-in voices; Drama 3 supports speaker and event tags |
| MuseScore General / GeneralUser GS / SGM-V2.01 | General MIDI sound fonts for score rendering; needs no inference engine or GPU, perfectly deterministic |
The samples were generated locally with yovoice via the CLI, using the models labeled next to each clip. 'Calm' and 'Happy' share the same text and reference voice, with only the emotion parameter adjusted. The website plays pre-rendered files and does not run models in your browser.
By default, they are stored in the local ~/.yovoice directory on your computer. You keep full control of your drafts, assets, and models. Uninstalling the app won't erase them, and you can easily back up or move this directory when changing machines.
Local generation and score rendering work completely offline.Network access is only required for installation, downloading models or sound fonts, and using optional cloud AI services. Once the required files are on your machine, speech and music generation run locally; score rendering needs neither an engine nor a GPU.
Yes, through an optional AI service.Connect DeepSeek, Qwen, Kimi, OpenAI, Anthropic, Ollama, or a compatible API, and the assistant can arrange entire pieces or fill in song styles, lyrics, and titles. Other features remain fully functional without it. Your API keys stay on your machine, and requests are sent only to the service you choose.
The current inference engine is audio.cpp v0.9.3. The desktop app supports:
For lightweight narration, 4 GB+ RAM is recommended; for high-fidelity cloning and voice design, 4–6 GB VRAM or 8 GB+ on Apple Silicon; for full songs and music generation, 12 GB+ VRAM or 16 GB+ on Apple Silicon. Actual usage depends on the model and precision.
Automatic audio health checks and broadcast target loudness exports are supported:Automatically flags clipping, clicks, cut-off endings, and long silences after generation, with hover details and timestamps. Export offers target loudness presets for audiobooks (−18 LUFS), podcasts (−16 LUFS), and video platforms (−14 LUFS), keeping true peaks strictly under −1 dBTP.
yovoice is licensed under Apache-2.0. Whether generated content can be used commercially depends on individual model licenses and your rights to the source materials:
Yes. The standalone CLI, Agent Skill, and MCP integrations can generate speech from text and reference voices, or run `yovoice score render` to produce mixes and stems, without opening the desktop app.
yovoice DESKTOP APPLICATIONLocal voice and music · Version history · Automatic audio check · Target loudness export
Click an icon to have an AI read yovoice materials and answer your questions.
Scan with WeChat to follow.
