CosyVoice.cpp
An unofficial C++ and GGML port of the CosyVoice text-to-speech inference pipeline, focused on CosyVoice3, with zero-shot, instruct and cross-lingual modes and streaming output.
CosyVoice.cpp is a community-maintained C++/GGML port of the Python CosyVoice text-to-speech inference pipeline, currently focused on CosyVoice3. It is not affiliated with the official CosyVoice team. It supports zero-shot, instruct and cross-lingual synthesis modes, in both synchronous and streaming output, and can reuse pre-encoded reference audio to skip the ONNX frontend.
The project provides a core C/C++ inference library, a CLI synthesis tool, an OpenAI Speech API-compatible server with an embedded web UI for model and speaker management, and a GGUF quantization tool. It runs on CPU, CUDA, Metal, Vulkan and SYCL backends, with per-backend test status tracked in the repository.
It is written in C++ and licensed under MIT.