yovoice is an open-source local voice-creation app for macOS and Windows that turns text into speech with on-device TTS models, offering voice cloning, emotion control, and project management; per its README, it requires no cloud API calls and incurs no per-character charges. It runs IndexTTS 2.0 and 2.5, VoxCPM2, OmniVoice, and Qwen3-TTS (Base 0.6B/1.7B, CustomVoice 1.7B, VoiceDesign 1.7B) through the audio.cpp inference engine, with resumable model downloads and GGUF import; a companion standalone CLI and Agent Skill run the same pipeline outside the app, and the CLI also supports Linux x64. IndexTTS 2.0/2.5 offer Chinese/English (2.5 adds Japanese, Spanish, Arabic) voice cloning with emotion control; VoxCPM2 adds text-guided voice design and transcript-assisted cloning with multilingual handling and 48 kHz output; OmniVoice offers attribute-based voice design and non-verbal sound tags; Qwen3-TTS CustomVoice ships nine built-in speakers, and VoiceDesign builds a voice from a text description with no reference audio. Hardware acceleration covers Metal on Apple Silicon and CPU/CUDA/experimental Vulkan on Windows, where the CPU engine ships bundled while CUDA downloads separately. macOS builds require macOS 14+ on Apple Silicon as a .dmg; Windows builds target Windows 10/11 x64 as a setup.exe and install the WebView2 Runtime if missing. Models download inside the app; projects, voices, settings, and generation history live under ~/.yovoice, which persists after uninstalling; the app checks for updates from its menu and in the background, installing them on restart. The desktop shells are a native Swift app on macOS and a WPF/WebView2 (.NET, C#) app on Windows, both hosting a React/TypeScript/Vite UI over a Go core (cmd/yovoice, cmd/yovoice-service) that also powers the CLI. yovoice's own code is Apache-2.0; bundled components keep their own terms, including audio.cpp (Apache-2.0) and a trimmed FFmpeg 8.1.2 build (LGPL-2.1-or-later) shipped as a separate executable. Model weights download separately, not bundled in installers: IndexTTS weights are governed by their own model license, VoxCPM2 and Qwen3-TTS 12Hz 1.7B Base weights are Apache-2.0, and OmniVoice's code is Apache-2.0 but its pretrained weights are CC-BY-NC and restricted to non-commercial use, per the model author.