Kyläkauppias
- A super-tiny speech-to-speech neural network.
Details
Trained from scratch (absolutely no base model) to act as a Finnish grocery store assistant.
Ask for one of three products (in the Finnish language):
- A packet of bread (leipäpaketti, paketti leipää, yksi leipä..)
- A packet of cheese (juustopaketti, paketti juustoa, juustoa..)
- A packet of ham (kinkkupaketti, paketti kinkkua, kinkkua..)
The model understands orders with quantities between 1 and 9.
Online demo
https://olli-moisio-kylakauppias-demo.hf.space
Local usage
python -m venv kylakauppias_venv
source kylakauppias_venv/bin/activate
pip install -r requirements.txt
python app.py
Local microphone use requires https, for that set share=True in app.py.
Training data
- ~ 12000 Finnish text samples created with local Qwen 3.6 35B-A3B instance
- ~ 12000 Finnish audio samples created from text with local Chatterbox-TTS instance
Random noise and empty spaces were injected to audio samples during training for a more robust model.