Arabic diacritization · runs entirely on your device

Tashkeel Compare

A 1.08 million parameter model vocalizes Arabic with no server and no GPU. Put it beside far larger systems, paste a whole document, and judge the marks yourself.

Our smallest: 1.08 M
Runs on: Your CPU
Network: None

Arabic text to vocalize

Paste a document, or upload a file

Upload .txt Chunk budget 400 chars

Reference text (fully vocalized gold)

Ignore mark order (compares readings, not spelling conventions)

Marks are stripped from the reference, every model re-vocalizes the bare text, and each output is scored against the original.

What this page does and does not claim

Our two models run live in your browser through ONNX Runtime Web. Their output matches the published Python packages - the same class at every tested position - and CI checks exactly that.
Mishkala and FLAN-T5 outputs on the built-in samples are precomputed. They cannot run on a static page, and they are labelled as such wherever they appear.
Leaderboard numbers are quoted from the published SadeedDiac-25 table, produced by that benchmark’s own letter-level evaluator. They are never mixed with ours.
Every number says which evaluator produced it. The Board’s measured here rows come from the benchmark’s own letter-level evaluator, scored on all 1200 rows. The Score tab runs a different, self-contained scorer inside this page — letter-level DER over marked letters, word-level WER — and says so on the tab. Nothing is placed in a column it does not belong to.