Serveurperso/OmniVoice-GGUF

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

variant	base	tokenizer	use case
F32	2.46 GB	734 MB	reference, debug, conversion
BF16	1.23 GB	373 MB	source faithful, max precision
Q8_0	656 MB	289 MB	recommended default
Q4_K_M	407 MB	252 MB	lowest VRAM

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.

value	target
CUDA0	NVIDIA GPU, fastest path on Ada / Blackwell
Vulkan0	Cross vendor GPU (AMD / Intel / NVIDIA)
Metal	Apple Silicon GPU
CPU	CPU fallback, x86 variant auto selected

Quantization policy

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

tensor	dtype across all variants
RVQ codebooks, fc, fc2, project_in / project_out	F32
Snake activation alpha	F32
Conv kernels with non alignable rows (K=7,3,1)	F16 (in Q* variants)

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 weights : OmniVoice by Xiaomi / k2-fsa, CC-BY-NC 4.0 Upstream code : 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

The upstream code is Apache 2.0 but the pre-trained weights are CC-BY-NC because of constraints from the training data. These GGUF files are conversions of those weights and carry the same non commercial terms.