Research · 2026 – present
Kurdia
Language technology for every Kurdish dialect, on ordinary hardware: dialect identification, writing checker, OCR, speech recognition, speech synthesis and a small local language model.
Most Kurdish language tools serve standard Sorani, sometimes Kurmanji, and assume a GPU. Kurdia is built for all of it: Sorani in Arabic script, Kurmanji in Latin script, and the Gorani group — Gorani, Hawrami, Zazaki and Southern Kurdish — on a laptop, a phone, a Raspberry Pi or a browser tab.
Six modules, one at a time
- Dialect classifier — which variety, sub-variety and script a text is in, with a confidence, code-switch spans and the option to answer “unsure”. Every other module asks it first.
- Writing checker — spelling, grammar and style that respect the writer’s own dialect.
- OCR — printed and scanned Kurdish in both scripts.
- Speech to text, with a spoken-dialect router.
- Text to speech per dialect.
- A small Kurdish language model that runs locally.
Principles
- Dialect first. No module pushes a writer into another dialect.
- Rules, learning and control together. Each variety’s grammar is written as a formal grammar drawn from reference grammars, with page citations. Supervised models handle what rules cannot, and reinforcement learning decides what evidence to compute next and when to stop.
- Runs anywhere. Shipped inference imports only NumPy and ONNX Runtime; large GPU models are teachers, never runtime.
- Publishable, and honest. Every module ends in a paper for a top-quartile venue. No accuracy figure appears without a named benchmark behind it; test splits are frozen and split by source.
The dialect classifier is in its first phase: scaffolding, data contracts and baselines.