Zyphra/ZONOS3-Mini
COMPACT TTS WITH VOICE CLONING
Multilingual text-to-speech model with zero-shot voice cloning. Runs locally on GPU, CPU, Apple Silicon, and in browser via WebGPU.
// KEY FEATURES
- Four backends: CUDA/SGLang, MLX, ggml/C++, WebGPU in browser
- Zero-shot voice cloning from reference audio or embedding
- Unified HTTP API and web studio for all backends
- OpenAI-compatible /v1/audio/speech endpoint
// DEEP DIVE: HOW IT WORKS
ZONOS3-Mini is an inference runtime for Zyphra's multilingual TTS model. The project provides four server implementations for different platforms: CUDA version on SGLang for NVIDIA GPU, MLX server for Apple Silicon, C++/ggml server for CPU, CUDA, Metal and Vulkan without Python at runtime, and an in-browser WebGPU build. All backends serve the same HTTP API and the Zonos Voice Studio web interface.
Inside is the ZONOS3-Mini model with zero-shot cloning support: generation can be conditioned on reference audio (base64) or a pre-saved speaker. There is a native /v2/generate endpoint with streaming and an OpenAI-compatible /v1/audio/speech. Weights (~9 GB for PyTorch, GGUF for others) are downloaded from Hugging Face without a token. The text normalizer comes as a Rust library (ggml) or Python wheel.
Requirements: uv and Python 3.12; for CUDA server — CUDA 12.8 toolkit and NVIDIA GPU (tested on H100), first start JIT-compiles kernels for several minutes; ggml server needs Rust toolchain and CMake; MLX — Apple Silicon only; in-browser — browser with WebGPU. Weights download on first run to HF cache.
// USE CASES
Best for: Developers of voice interfaces needing local TTS with cloning
- Deploy a local TTS server with voice cloning on your GPU or Mac
- Integrate speech generation into an app via OpenAI-compatible API
- Run voice synthesis in the user's browser without a backend
// QUICK START
git clone --recurse-submodules git@github.com:Zyphra/ZONOS3.git
cd ZONOS3
uv sync --locked --all-packages # Linux with an NVIDIA GPU
uv sync --locked --package zonos3-mlx # Apple SiliconCommands are taken from the repository README — check the current version on GitHub before running.
New repositories every 30 minutes
REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.
Join on Telegram