OmniVoice GGUF

OmniVoice GGUF

Provides GGUF weights of the OmniVoice text-to-speech model for the omnivoice.cpp runtime, with voice design and voice cloning examples. For developers who want local, multilingual speech output.

Made by Serveurperso on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository holds GGUF weights of OmniVoice, a text-to-speech model, for use with omnivoice.cpp, a C++17 and GGML port of the original k2-fsa project. The card says it speaks 646 languages without per-voice training and writes 24 kHz mono audio. Two GGUF files load together. A base file with a Qwen3 0.6B backbone turns text into tokens, and a tokenizer file turns tokens back into audio.

Who it suits: It suits developers who want speech output that runs on their own hardware, and researchers who test zero-shot voice design and cloning. The weights are non-commercial, so it fits personal, study and research use.

Industry | Who uses it | How they use it | Example
Software | App developers | Add on-device speech output through omnivoice.cpp | Reading text from an app aloud as a 24 kHz wav file
Media | Hobby video and podcast makers | Generate narration for non-commercial projects | Producing a voice-over track for a personal video
Research | Speech researchers | Try zero-shot voice design and voice cloning | Running the card's two example scripts
Education | Students of speech technology | Compare quantized variants of one model | Loading the Q8_0 and Q4_K_M files and listening to both

If you need commercial use, you can skip it, because the weights carry non-commercial terms.

Where to use it:

• omnivoice.cpp running on CPU, CUDA, ROCm, Metal or Vulkan
• A terminal for building the project and running its example scripts
• Windows, macOS or Linux with Python, for downloading the files from the hub

Value | Target
CUDA0 | NVIDIA GPU, the fastest path on Ada and Blackwell cards
Vulkan0 | GPUs from AMD, Intel or NVIDIA
Metal | Apple Silicon GPU
CPU | Fallback, with the x86 variant picked automatically

Use case | Who | What they get back
Voice design | Developers and makers | A wav file named tts.wav from the example script
Voice cloning | Researchers | A wav file named clone.wav from the example script
Low-memory use | People with small GPUs | The Q4_K_M files, which the card lists as the lowest VRAM option

How to set it up:

• Git, to clone the omnivoice.cpp project with its submodules
• A build setup for C++, since omnivoice.cpp is built from source
• The Hugging Face command line tool for downloading the two GGUF files
• Disk space for the pair you pick, such as 656 MB and 289 MB for the Q8_0 variant

1. Clone omnivoice.cpp from its GitHub page with submodules. The card gives the clone command, which holds a web address, so copy it from there.
2. Build it, here with the CUDA build script.
3. Make a models folder and download the base and tokenizer files into it.
4. Run the example scripts from the examples folder.

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

Variant | Base file | Tokenizer file | Use case
F32 | 2.46 GB | 734 MB | Reference, debugging and conversion
BF16 | 1.23 GB | 373 MB | Faithful to the source, with maximum precision
Q8_0 | 656 MB | 289 MB | The recommended default
Q4_K_M | 407 MB | 252 MB | The lowest VRAM

How to use it: Run tts.sh from the examples folder for voice design, and clone.sh for voice cloning. Set GGML_BACKEND to force a device. If you leave it unset, the runtime picks the best one available.

Tool | What it does
tts.sh | Designs a voice from text and writes tts.wav
clone.sh | Clones a voice and writes clone.wav
GGML_BACKEND | Forces one device, such as CUDA0, Vulkan0, Metal or CPU

Example: You run tts.sh from the examples folder after building and downloading the Q8_0 files. A speech file named tts.wav appears. Run clone.sh for cloning and you get clone.wav. Clone only voices you have the right to use.

For newcomers and experts: Text to speech means a program reads written text aloud. Zero-shot means it can produce a voice without being trained on that voice first. A good first try is the Q8_0 pair, which the card calls the default, followed by tts.sh.

Under the hood, the base model has a hidden size of 1024 and all its dimensions divide by 256, so it follows the standard llama.cpp K-quant in every variant. The tokenizer files combine HuBERT, DAC and RVQ and are not uniformly quantized. The RVQ codebooks, the Snake activation alpha and a few projection layers stay in F32, and conv kernels with rows that cannot align drop to F16 in the quantized variants.

Text goes in
The base GGUF turns it into tokens
The tokenizer GGUF turns tokens into audio
A 24 kHz mono wav file comes out

Good to know:

• The upstream weights are CC BY-NC 4.0 and the upstream code is Apache 2.0, because of limits on the training data.
• The audio codec is Higgs Audio v2 under Apache 2.0, and the omnivoice.cpp tooling is MIT.
• These GGUF files are conversions of the non-commercial weights and carry the same terms.
• License: CC BY-NC 4.0 for the weights, which does not allow commercial use
• Cost to run: none named
• Data it reaches: the models folder and the wav files the scripts write, and the source names no account
• Not for: commercial use, and cloning voices you have no right to use

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: gguf
• Parameters: 613M
• Weights: about 306 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 109,605
• Likes: 91
• License: cc-by-nc-4.0
• Languages: en, fr, de, es, it, pt, zh, ja
• Based on: k2-fsa/OmniVoice
• Tags: gguf, text-to-speech, voice-cloning, voice-design, ggml, omnivoice

Links

• Model: https://huggingface.co/Serveurperso/OmniVoice-GGUF
• Files: https://huggingface.co/Serveurperso/OmniVoice-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Serveurperso/OmniVoice-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.