Lekhani

Type Hindi in Latin letters and get Devanagari. The model runs in this page, on your own device, so nothing you type is sent anywhere.

How good it is

Aksharantar Hindi test (AI4Bharat): 9,991 romanised words, top-1 exact match, the same words and the same scorer for every system. Measured by us on 28 October 2026.

System	What you download or hold	Words exactly right
Lekhani	0.87 MB in all	65.9%
IndicXlit, model on its own	11.5M parameters, 14× Lekhani's	58.9%
IndicXlit, full Python package	the model and a 4.2M-word list, 358 MB in memory (about 400× all of Lekhani)	68.7%

Lekhani is 7.0 points ahead of IndicXlit's model on its own. IndicXlit's full package is still 2.8 points more accurate, if you can hold its word list.

Where it falls short

• It reads one word at a time, so main (I) and mein (in) both become में. Tap the suggestion to fix it.
• Some common words come out wrong: acha becomes अच् (accha works), kal becomes काल (कल is its second choice).
• Hindi only, Latin letters to Devanagari only. About one word in three on the hard test above is wrong at top-1.

Use it in your own site or app

Five files, 0.87 MB, no dependencies: translit.js, model.q4.gz, model.json.gz, words.bin.gz, rerank.json. Apache-2.0 for the model and the runner, CC0 for the word list.