LeonardoDaviti/Georgian-NeMo-GGUF

Georgian-NeMo GGUF (transcribe.cpp)

GGUF build of EchoesML/Georgian-NeMo for transcribe.cpp, the engine inside Handy. Published with the permission of EchoesML.

• Architecture: FastConformer hybrid RNN-T + CTC, 115M parameters, 16 kHz mono input
• Origin: EchoesML fine-tuned it from nvidia/stt_ka_fastconformer_hybrid_large_pc (CC-BY-4.0)
• Decoding: the GGUF uses the RNN-T head with greedy decoding

File	Quant	Size
georgian-nemo-Q8_0.gguf	Q8_0	135 MB

Accuracy check

The conversion was checked against the original NeMo PyTorch model on 40 FLEURS ka test clips (seed-0 sample, Colab):

Model	WER %	Agreement with NeMo
NeMo (PyTorch)	13.72	—
GGUF Q8_0	13.72	100%

On an AMD Radeon 8060S with Vulkan it runs at roughly 190–200× real time.

Use

Handy (0.9+): put the .gguf file directly in Handy's models folder (not in a subfolder), then restart Handy.

OS	Handy models folder
Linux	~/.local/share/com.pais.handy/models/
macOS	~/Library/Application Support/com.pais.handy/models/
Windows	%APPDATA%\com.pais.handy\models\

On Linux/Wayland, set the paste method to Ctrl+V (clipboard). Typing through ydotool drops Georgian letters.

Python:

# pip install transcribe-cpp transcribe-cpp-native
from transcribe_cpp import Model
model = Model("georgian-nemo-Q8_0.gguf", backend="vulkan")  # or "cpu"
print(model.session().run(pcm_16k_mono_float32, language="ka").text)

Credits

• Acoustic model and fine-tune: EchoesML
• Base model: NVIDIA (CC-BY-4.0)
• GGUF conversion: transcribe.cpp convert-parakeet.py, using a custom georgian-nemo variant profile