Breeze-TTS-2.cpp
A C++ and GGUF reimplementation of the Breeze-TTS-2 bilingual English and Mandarin text-to-speech model, supporting voice design, cloning, direction, and experimental voice conversion with a Vulkan backend.
Breeze-TTS-2.cpp reimplements the Breeze-TTS-2 speech model in C++ on top of ggml and GGUF, running on NVIDIA, AMD, and Intel GPUs through Vulkan with a CPU fallback. It generates 24 kHz bilingual English and Mandarin speech at roughly 1.2 times real time on a mid-range GPU.
Four modes are supported: voice design from a text description, cloning a voice from a reference clip and its transcript, directing an existing voice's tone or pace with an instruction, and experimentally converting a recording into a different voice. The project ships a command-line interface, a streaming HTTP and WebSocket server with a web UI, and a plain C shared library for language bindings, and it can convert or download prebuilt GGUF model weights.