Breeze-TTS-2-GGUF

Derived from Breeze TTS 2 by BreezeBlue and licensed for research and non-commercial use only.

Unofficial GGUF distribution of Breeze TTS 2 for audio.cpp. This repository provides the BF16 baseline and F16, Q8_0, Q6_K, and Q4_K conversions. It is not affiliated with or endorsed by BreezeBlue or the audio.cpp maintainers.

Provenance and modifications

• Original model: BreezeBlue/Breeze-TTS-2.
• Conversion baseline: audio-cpp/audio.cpp-gguf, Breeze-TTS-2-GGUF/breeze-tts-2-bf16.gguf.
• Conversion tool: audiocpp_gguf.exe from audio.cpp 0.9.2, build 0097d365 (2026-10-15).
• Changes: storage conversion or quantization of eligible weights to the format named in each filename, preserving the packaged model resources. No additional training or fine-tuning was performed.
• BF16 is the baseline package; F16 and quantized variants are generated from it. Tensor precision can be mixed: the filename describes the requested conversion type, not a guarantee that every tensor uses that type.

Files

File	Bytes	Format
breeze-tts-2-bf16.gguf	7,342,918,784	BF16 baseline
breeze-tts-2-f16.gguf	7,342,918,752	F16
breeze-tts-2-q8_0.gguf	4,933,787,232	Q8_0
breeze-tts-2-q6_k.gguf	4,856,941,728	Q6_K
breeze-tts-2-q4_k.gguf	4,313,648,736	Q4_K

The models support English and Mandarin speech generation, reference-audio voice cloning, and voice instructions through audio.cpp. Reference audio requires an accurate transcript.

Compatibility

These files use audio.cpp's Breeze package layout. Sharing the GGUF container format does not establish compatibility with other inference engines, including HoppouAI/Breeze-TTS-2.cpp or llama.cpp.

Use a compatible audio.cpp release with the breeze_tts family. The packages include supporting resources; no Python inference environment is required. Low-bit quantization quality and backend compatibility should be evaluated for the intended workload.

Usage

After downloading one file, run audio.cpp's CLI. For example, on Windows with a CUDA build:

.\audiocpp_cli.exe `
    --task tts `
    --family breeze_tts `
    --model '.\breeze-tts-2-q6_k.gguf' `
    --backend cuda `
    --text 'Hello. This speech is generated locally.' `
    --out '.\speech.wav'

On an Apple Silicon Mac, use a Metal-enabled audio.cpp build, its audiocpp_cli executable, and --backend metal. See the Breeze model documentation for cloning, streaming, and Metal precision settings.

Evaluation and limitations

No formal quality benchmark or complete matrix of quantization/backend tests is reported here. Conversion success is not a quality measurement. Compare pronunciation, speaker similarity, long-form stability, and streaming behavior against the BF16 baseline on the hardware you intend to use. No macOS real-time performance claim is made.

License and attribution

The model and these derivatives remain subject to the BreezeBlue Research and Non-Commercial License Agreement. Quantization does not remove its restrictions or grant commercial rights. Preserve this license and the accompanying NOTICE when redistributing. Consult the original license for the complete terms and use restrictions.