HoppouAI's Breeze-TTS-2 Hits 860K Downloads With No Python Required
A community C++ port brings BreezeBlue's 3B bilingual TTS model to any GPU through Vulkan, with voice cloning and sub-realtime generation on consumer hardware.
Takeaways
- HoppouAI released Breeze-TTS-2.cpp, GGUF quants plus a C++ runtime for BreezeBlue's 3B TTS model.
- Runs on NVIDIA, AMD, Intel GPUs via Vulkan, with CPU fallback and no Python required.
- Hits roughly 1.2x realtime at Q8_0 on an RTX 3060, 24 kHz bilingual English and Mandarin output.
- Supports voice cloning, text-described voice design, and experimental voice conversion with timing preserved.
- Experimental -dd quants shrink the depth decoder but degrade past 45 seconds of continuous generation.
- Weights are research and non-commercial only under the BreezeBlue license, runtime code is Apache 2.0.
HoppouAI’s GGUF build of BreezeBlue’s Breeze-TTS-2 reached nearly 860,000 downloads on Hugging Face. It combines quantized weights with Breeze-TTS-2.cpp, a C++ inference runtime that supports NVIDIA, AMD, and Intel GPUs through Vulkan, plus a CPU fallback. Developers can self-host the voice model without a Python runtime or CUDA-specific toolchain.
A four-stage speech stack
BreezeBlue’s original model is a 3-billion-parameter, open-weight text-to-speech system for English and Chinese. It accepts natural-language instructions and supports three generation modes:
| Mode | Inputs | Result |
|---|---|---|
| Voice cloning | A reference clip and its exact transcript | Speech that reproduces the reference voice’s characteristics |
| Voice design | A natural-language voice description | A new voice generated without reference audio |
| Voice direction | Reference audio, its transcript, and delivery instructions | A cloned voice with controlled tone, emotion, pace, and delivery |
The C++ project reimplements inference around unchanged model weights. Its runtime covers the full pipeline: a T5Gemma2 encoder represents the text and instructions, a Qwen3 backbone predicts the main token sequence, a depth decoder fills the acoustic codec layers through 15 serial steps, and a vocoder renders the final waveform. The GGUF files are tailored to this runtime, so llama.cpp cannot load them, and the original PyTorch checkpoints cannot replace them directly.