How we test accuracy
Every number on this page comes from our test run, written to a results file and read by the site build. No number here is typed by hand. Last measured: 6 Nov 2026.
What we test today
We test in two ways. Against other converters, on words held out from building our tables: that shows our tables match the layout other converters use. And on text from real government documents typed in Kruti Dev 010: that shows what happens with real typing. For real documents, the result that counts is the one on documents that played no part in building or fixing the converter (held out).
Kruti Dev 010 results
| Test | Result | What it means |
|---|---|---|
| Kruti Dev → Unicode | 980 / 980 | 1,000 held-out Hindi words were written in Kruti Dev by another converter. We read them back and compared each with the original word. 20 words were left out because that converter dropped the nukta dot, so its Kruti Dev text no longer held the letter. Raw result before leaving them out: 980 / 1,000. |
| Unicode → Kruti Dev | 1,000 / 1,000 | Another converter read our Kruti Dev output back. Where it disagreed, a second converter decided. Raw agreement with the first converter: 995 / 1,000; the differences were the reph on a joined letter (as in पार्श्व) and words that converter misreads itself. |
| Same codes as another converter | 921 / 1,000 | Not an error count. The differences are other spellings that draw the same letters, for example प्र written iz instead of a ready-made glyph, or a nukta kept instead of dropped. |
| Round trip (self-check) | 8,315 / 8,322 | Unique words converted to Kruti Dev and back to Unicode. This checks our own consistency, not accuracy. The words left are malformed source words or letters with no Kruti Dev glyph. |
Real Kruti Dev 010 documents: held out
On 6 Nov 2026 we converted the Kruti Dev 010 text of 3 government PDFs that were never used to find or fix errors (907 words): a gazette notification with a land schedule and two state notices. Kruti Dev 010 met our 98% bar on them: 99.2% of 907 words were right after the hand check described below (96.0% were on the word list before it).
The sample is small. With 907 words, the likely range for this kind of text is 98.4% to 99.6% (95% interval). Three documents of two kinds cannot speak for every office, and only one word in them used Word's curly quote marks, which we read wrongly: the text was too short to tell ष from श, so the converter now says so and lets you choose. We will add longer documents to this test.
Real Kruti Dev 010 documents: the documents we fixed against
On 6 Nov 2026 we converted the Kruti Dev 010 text of 2 real government PDFs (9,517 words). Such documents have no answer key, so a word counts as right when it is a real Hindi word: each converted word is looked up in a large list of Hindi word forms made from Hindi Wikipedia, and every word not on the list is checked by hand against the printed page. Real names and government terms missing from the list are then counted as right; everything else stays wrong.
| Version | Words on the list | After the hand check |
|---|---|---|
| Before the fixes | 95.5% | 96.6% |
| After the fixes | 98.4% | 99.5% |
These results are measured on the documents the fixes were made for, so they are not a fair test on their own. These documents showed us the errors we then fixed: Word's curly quote marks standing for श or ष, ॉ typed as ा and ॅ, a bindu typed before a vowel sign, े typed before ा, the visarga key used as a colon, and signs typed twice. The held-out result above is the one to go by.
The words still counted as wrong after the hand check were typing errors in the source documents (a sign typed twice, a letter left out or swapped) or words run together in the PDF; the converter copies those as they are. A wrong conversion that happens to make another real word would not show up in this test; a check by eye of the opening part of one document found none.
The test words
The words come from Hindi Wikipedia articles (49,129 running words, 8,322 different words). They are test data only and are not published here. We drew two random samples of 1,000 words. The first was used to find and fix one missing glyph, so it no longer counts as a fair test; the results above are from the second sample, which was not used to change anything.
What is not tested yet
- Longer held-out Kruti Dev 010 documents, especially ones with Word's curly quote marks (planned next).
- Other fonts. DevLys 010, Chanakya and Preeti get their own tests before their pages open.
- Word (.docx) files, which the site does not open yet.
- The rare codes listed as known limits on the Kruti Dev to Unicode page.
Found a wrong word?
Email hello@bhashakit.com with the word, the page and what you expected. Every report is checked against our table, and fixes are tested again before release.