tiny-finglish

How accurate is it?

Each setup converts a set of real sentences, and we count how many words come out the way a person wrote them. Spellings Persian writers disagree on count as right: می‌کنم, میکنم and می کنم are the same word, and so are زمانیکه and زمانی که.

What's inside

A small character-level network and a rule engine, both running in plain JavaScript in this tab. No server, no API key.

What you download

Code: rules, tokenizer, decoder	21.4 KiB
Word frequencies and vowels	68.6 KiB
The model's weights	84.4 KiB
Sentence context (optional)	73.6 KiB

Compressed sizes. The best setup, both engines together, is 174 KiB, about the size of one photo.

How a sentence is converted

1. Links, emails, numbers, code and English words are set aside and kept exactly as typed.
2. Each remaining word gets a handful of possible Persian spellings, from the rules and from the model.
3. The spellings are ranked by how well they fit what was typed and how common they are.
4. A last pass looks at position: a word that ends a sentence may take the ه of خوبه.

Speed

A keystroke re-converts only the word being typed, in well under a millisecond. A whole new sentence takes about 35 ms the first time.

Source, measurements and known limits are in the project README.