Kizagan-TTS-v1.0-GGUF

Kizagan-TTS-v1.0 — GGUF weights

These are GGUF conversions of AlicanKiraz0/Kizagan-TTS-v1.0 for the llama.cpp-omni VoxCPM2 runtime. The model was previously released under the name Kahya-TTS; the weights are identical and the release has simply been renamed. The earlier Kahya-* GGUF files in this repository have been replaced by the Kizagan-* files below.

The Acoustic file is required for every call, paired with either BaseLM variant. Only the BaseLM is quantized; the acoustic path stays F16 for quality.

Model and training provenance

Getting Started

1. Clone

git clone https://github.com/tc-mb/llama.cpp-omni.git
cd llama.cpp-omni

2. Build

# macOS (Metal — auto-detected):
cmake -B build-omni -DCMAKE_BUILD_TYPE=Release
cmake --build build-omni --target voxcpm2-cli llama-quantize llama-tts-server -j

# Linux / Windows (CUDA):
cmake -B build-omni -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON
cmake --build build-omni --target voxcpm2-cli llama-quantize llama-tts-server -j

# Linux / Windows (Vulkan):
cmake -B build-omni -DCMAKE_BUILD_TYPE=Release -DGGML_VULKAN=ON
cmake --build build-omni --target voxcpm2-cli llama-quantize llama-tts-server -j

# CPU only:
cmake -B build-omni -DCMAKE_BUILD_TYPE=Release
cmake --build build-omni --target voxcpm2-cli llama-quantize llama-tts-server -j

Usage

Basic TTS (GPU by default; add --cpu to force CPU)

# Q8_0 BaseLM (recommended):
voxcpm2-cli \
    -t "Merhaba, ben Alican Kiraz tarafından oluşturulan Kizagan Türkçe TTS modeliyim." \
    -o output.wav \
    --timesteps 16 --cfg 2.0 --seed 42 \
    Kizagan-BaseLM-Q8_0.gguf Kizagan-Acoustic-F16.gguf

# F16 BaseLM (maximum accuracy / reference):
voxcpm2-cli \
    -t "Merhaba, ben Alican Kiraz tarafından oluşturulan Kizagan Türkçe TTS modeliyim." \
    -o output.wav \
    --timesteps 16 --cfg 2.0 --seed 42 \
    Kizagan-BaseLM-F16.gguf Kizagan-Acoustic-F16.gguf

Recommended parameters for this fine-tune: --timesteps 16 --cfg 2.0 --seed 42 (fixed seed → reproducible output).

Voice cloning (3–7 s clean mono PCM reference)

# References must be PCM (fmt=1) WAV; convert MP3/float-WAV with:
ffmpeg -i reference.m4a -ac 1 -ar 48000 -c:a pcm_s16le ref-pcm.wav

voxcpm2-cli \
    -t "Klonlanmış sesle yeni bir cümle okuyorum." \
    -r ref-pcm.wav \
    -o clone.wav \
    --timesteps 16 --cfg 2.0 --seed 42 \
    Kizagan-BaseLM-Q8_0.gguf Kizagan-Acoustic-F16.gguf

You can also design a voice without a reference by prefixing the text, e.g. -t "(Genç bir kadın, yumuşak ve sıcak ses)Merhaba dünya".

HTTP server

llama-tts-server --host 127.0.0.1 --port 8222 \
    --voxcpm2-base-lm Kizagan-BaseLM-Q8_0.gguf \
    --voxcpm2-acoustic Kizagan-Acoustic-F16.gguf

curl http://127.0.0.1:8222/v1/audio/speech \
    -H "Content-Type: application/json" \
    -d '{"model":"voxcpm","input":"Sunucudan gelen Türkçe ses testi.","cfg_value":2.0,"inference_timesteps":16,"seed":42}' \
    --output output.wav

Conversion

python tools/omni/voxcpm2/convert_voxcpm2_to_gguf.py \
    --model /path/to/Kizagan-TTS/model.safetensors \
    --vae /path/to/Kizagan-TTS/audiovae.pth \
    --config /path/to/Kizagan-TTS/config.json \
    --tokenizer-dir /path/to/Kizagan-TTS \
    --output ./gguf --dtype f16

llama-quantize ./gguf/VoxCPM2-BaseLM-F16.gguf ./gguf/Kizagan-BaseLM-Q8_0.gguf Q8_0

Acknowledgments and attribution

A heartfelt thank you to the OpenBMB team and the VoxCPM contributors for making VoxCPM2, its weights, training tools and implementation available. Kizagan builds directly on their work; this Turkish adaptation and its evaluation would not have been possible without that foundation. OpenBMB ekibine ve VoxCPM2'ye emek veren tüm araştırmacılara çok teşekkür ederiz.

• Base model: openbmb/VoxCPM2
• Upstream implementation: OpenBMB/VoxCPM
• Technical report: Zhou et al., VoxCPM2 Technical Report (2026)
• Earlier work: VoxCPM (2025)

This is an independent community adaptation; OpenBMB authorship applies to the foundation model and upstream implementation. Licensing and attribution of the original release apply; see the model repository's LICENSE and NOTICE.