# yovoice — complete public guide

> yovoice v0.5.0 is an open-source, local-first audio and music creation studio for macOS, Windows, and Linux.

## Product definition

yovoice creates expressive speech, cloned voices, full vocal songs, instrumentals, sound effects, and multi-track scores. It provides a desktop app, a standalone CLI, an Agent Skill, and HTTP API / MCP integrations. The website's audio samples are pre-generated speech files; the browser does not run a model.

## Positioning and comparison with alternatives

- **Versus cloud platforms (ElevenLabs, Suno)**: 100% offline and local-first. No character limits, no monthly subscriptions, and absolute privacy for sensitive manuscripts or proprietary reference voices.
- **Versus open-source voice tools (Voicebox, VoiceStudio, Applio)**: While Voicebox and VoiceStudio focus primarily on speech generation and cloning, yovoice unifies voice cloning, multi-role story dubbing, AI song creation with lyrics, and multi-track MIDI score arranging in one desktop application.
- **Versus raw script repositories (GPT-SoVITS, CosyVoice, F5-TTS)**: yovoice relies on native C++ execution (audio.cpp v0.9.3) with pre-bundled CUDA (Windows) and Metal (macOS) runtimes. No Conda, PyTorch, or Python dependencies are required.
- **Agent and MCP ecosystem**: Native Model Context Protocol (MCP) server support allows Claude Code, Cursor, and other agentic coding assistants to control audio and music workflows programmatically.

## Project types and creative workflow

- **Story dubbing and speech generation**: Import text, EPUB, or subtitle files, split them into lines, assign characters, generate line by line, and edit clips on a timeline. Includes automatic post-generation audio health checks and target loudness export presets (−18, −16, −14 LUFS).
- **Voice and expression**: Use a reference voice, match a reference performance, adjust emotions, blend voices, or describe the delivery in words. VoxCPM2, OmniVoice, Qwen3-TTS, Breeze TTS 2, Ming 3.0 TTS, XTTS v2, Voxtral, and Drama 3 provide diverse voice-design and cloning workflows.
- **Music and sound effects**: Describe a style and write section-marked lyrics. ACE-Step 1.5 Turbo / XL Turbo creates songs or instrumentals; YuE2 3B creates vocal songs and score-based instrumentals; Stable Audio 3 creates instrumental music (up to 6 minutes) and sound effects (1–30 s) with rework and inpainting. The playback page synchronizes lyrics word by word.
- **Score**: Write each track by bar and beat, import MIDI, or let an Agent compose. General MIDI sound fonts render each track, role-based levels shape the mix, and the result can be exported as MIDI, score JSON, a mixed WAV, or stems. The same score always renders the same way.
- **Takes and editing**: Generated audio is kept as versions. The workspace supports previewing, timeline editing, and export.

## Models and sound fonts

Supported speech and music model families are IndexTTS 2.0 / 2.5, VoxCPM2, OmniVoice, Qwen3-TTS Base / CustomVoice / VoiceDesign, Breeze TTS 2, Ming 3.0 TTS, Voxtral 4B TTS, Plapre 3, XTTS v2, Gradium Phonon, Drama 3 / Drama3-small, Kokoro-82M 1.0 Official / 1.1-zh, ACE-Step 1.5 Turbo / XL Turbo, Stable Audio 3 (Small Music, Small SFX, Medium), and YuE2 3B.

Kokoro-82M 1.0 Official provides 49 built-in voices and supports voice blending. Qwen3-TTS CustomVoice provides 9 built-in voices. Ming 3.0 TTS supports 7-language voice design and cloning. XTTS v2 supports 17-language voice cloning. Voxtral provides 20 built-in voices across 9 languages. Drama 3 supports speaker and event tags. ACE-Step 1.5 Turbo supports 50+ vocal languages.

Score projects use MuseScore General, GeneralUser GS, or SGM-V2.01. Score rendering needs a sound font but no inference engine or GPU.

See [models and licenses](https://yovoice.leemysw.com/models.md) for the complete parameter, precision, file-size, voice-count, and hardware tables.

## Optional AI services and data flow

DeepSeek, Qwen, Kimi, OpenAI, Anthropic, Ollama, and compatible APIs can optionally write a whole score or fill in a song's style, lyrics, and title. Everything else works without an AI service.

Keys are stored only on the computer. When an AI feature is used, the request and the fixed score or lyrics instruction go directly to the selected service. Local model inference does not require these services.

## Platforms and hardware

- **macOS**: macOS 14 or later on Apple Silicon, with Metal acceleration.
- **Windows**: Windows 10/11 x64, with CPU, NVIDIA CUDA, and experimental Vulkan support. Releases include CUDA 12.4 and 13.3 installers.
- **Linux**: Linux x64 (Debian/Ubuntu, .deb and .tar.gz with install.sh), with bundled CPU engine and optional Vulkan GPU engine.
- **Engine**: audio.cpp v0.9.3. Ming 3.0, Voxtral, XTTS v2, and Drama 3 require this engine version.
- **CLI platforms**: macOS Apple Silicon, Windows x64, and Linux x64.