fablevi / OmniVoice-GGUF-clone

OmniVoice GGUF

GGUF weights for omnivoice.cpp , a C++17/GGML port of OmniVoice (k2-fsa/OmniVoice). Multilingual zero shot TTS, 646 languages, 24 kHz mono. Runs on CPU, CUDA, ROCm, Metal, Vulkan.

Files

Two GGUFs load together :

omnivoice-base-{variant}.gguf Qwen3 0.6B backbone, text -> tokens omnivoice-tokenizer-{variant}.gguf HuBERT + DAC + RVQ, tokens <-> 24 kHz audio

Quick start

git clone --recurse-submodules https://github.com/ServeurpersoCom/omnivoice.cpp.git
cd omnivoice.cpp && ./buildcuda.sh

mkdir -p models
huggingface-cli download Serveurperso/OmniVoice-GGUF \
    omnivoice-base-Q8_0.gguf omnivoice-tokenizer-Q8_0.gguf \
    --local-dir models

cd examples
./tts.sh        # voice design   -> tts.wav
./clone.sh      # voice cloning  -> clone.wav

Backends

Set GGML_BACKEND to force a device, otherwise the runtime picks the best one available.

Quantization policy

Tokenizer GGUFs are not uniform quants. Three categories get a dedicated treatment :

Same fallback as llama.cpp tensor_type_fallback : F16 has no block size and matches the runtime target dtype on every backend. The base LM (Qwen3 0.6B, hidden = 1024) has all dimensions divisible by 256 so the fallback never triggers, the LM follows standard llama.cpp K-quant across variants.

License

Upstream model : OmniVoice by Xiaomi / k2-fsa, Apache 2.0 Audio codec : Higgs Audio v2 (bosonai/higgs-audio-v2-tokenizer), Apache 2.0 GGUF tooling : omnivoice.cpp, MIT