ZONOS3-Mini: Real-time voice cloning in 43 languages with 100M parameters

Zyphra's ZONOS3-Mini is an open-source text-to-speech model that performs natural speech synthesis and high-resolution voice cloning with 100 million active parameters. It runs in real time on smartphones and laptops, and can be executed directly in browser environments without server installation via WebGPU.

1 / 4

With just 5–10 seconds of audio samples, it can replicate a specific speaker's voice or design characters by adjusting age, accent, and timbre. By controlling emotion, speed, and delivery style at the word level, it can be applied to various scenarios such as in-game voices, narration, and voice assistants.

It supports 43 languages, including English, Korean, and Japanese. In an H100 80GB environment with 8 concurrent requests, it achieves a time to first audio of 80ms and processes 457 seconds of audio per second. It is released under the Apache 2.0 license, allowing for commercial use, and can be served as an OpenAI-compatible API via SGLang-Omni.

Various native runtimes are provided, including a CUDA server for NVIDIA GPUs, an MLX server for Apple Silicon, and a C++/ggml server supporting CPU and Vulkan. Developers can test the model directly in the browser using Zonos Voice Studio or build local inference pipelines by integrating with existing infrastructure such as llama.cpp and Ollama.