Research
The Hindi braille table most software ships is a revision out of date
Bharati Braille 2.1 was released on 4 January 2026. liblouis updated its Devanagari table to match on 26 February 2026. The table in the released version is still the 2014 one, and it writes 250 of the 5,000 commonest Hindi words differently.
This is not a bug in liblouis. The fix exists, it is correct, and it landed upstream a month after the standard moved. It simply has not reached a release yet, which means it has not reached the screen readers, embossing pipelines and translation tools that link against the released version — which is to say, effectively all of them.
That makes it worth publishing for a reason the six table defects are not. Those are rows to report. This is a thing to check before you conclude anyone is wrong: if your software and this site disagree about how to write फ, neither is broken, and the question is which revision each of you is following.
What changed
| Character | Was | Now | Why it matters |
|---|---|---|---|
| फ | 2-3-5 | 1-2-4 | The 2014 table gives फ dots 2-3-5; Bharati 2.1 gives it 1-2-4. फ is one of the commonest consonants in Hindi, so this change alone alters 156 of the five thousand commonest words — फिर, साफ, फोन, तरफ. |
| क़ | 1-3 | 5 then 1-3 | The 2014 table deleted the nukta outright, so क़ and क, ज़ and ज, फ़ and फ were the same cells — a distinction that changes the word disappeared. Bharati 2.1 gives the nukta a cell of its own, dot 5, written in front of the letter. |
| ऩ | 5-6 | 5 then 1-3-4-5 | In the 2014 table ऩ shared its cell with anusvara, ऱ with ड़ and ऴ with ल — three letters a reader could not tell from three other things. Bharati 2.1 removes all three by writing them with the nukta prefix. |
| १ | 1 | 3-4-5-6 then 1 | The 2014 table wrote no number sign before Devanagari digits, so १ and अ were the same single cell, while 1 written in Latin digits did get the sign. Bharati 2.1 writes the number sign for both. |
The single largest cause is फ, one of the commonest consonants in Hindi. Dots 2-3-5 in the old table, dots 1-2-4 in Bharati 2.1 — and that one change alone accounts for 156 of the 250 differing words.
The nukta change is the one with the worst failure mode. The 2014 table deleted the mark outright, so क़ and क, ज़ and ज, फ़ and फ were the same braille. A distinction that changes the word simply was not written — and nothing on the page shows a reader that something is missing.
How this was verified without the published document
There was no network path to the standard itself from where this was built, so the check had to come from somewhere else — and Bharati's own design supplies it.
Bharati Braille is one code across many Indian scripts. The cell that is क in Devanagari is the same sound in Bengali, Gujarati, Kannada, Malayalam, Oriya and Telugu, and liblouis carries a separate hand-written table for each. Those tables are therefore several independent statements of the same code, and they can be compared with each other.
Comparing Devanagari against 6 sibling scripts, character by character at the same position in each Unicode block:
| Characters | |
|---|---|
| Where every table that defines the character agrees | 76 |
| Where Devanagari stands alone against its siblings | 15 |
Five of those 15 — ऒ, ऴ, ऩ, ऱ and the vocalic vowels — turned out to be exactly the characters Bharati 2.1 changed. The comparison found the defects before the corrected table was found, which is the strongest evidence available here that it is the corrected one.
How to check your own tools
Put फिर through whatever writes your Hindi braille and look at the first cell. Dots 2-3-5 means the 2014 table; dots 1-2-4 means Bharati 2.1. The same test on क़लम answers the nukta question separately — an old table writes it identically to कलम, with no cell for the nukta at all.
What this site does
Every Hindi page here is written to Bharati 2.1, and the reference corpus is measured against the revised table rather than against whichever one happens to be installed. The 250 differing rows are committed in the repository, so the figure above is a count of lines somebody can open rather than a number to take on trust.
The Hindi section is at /hi/, the full measurement at /hi/research/, and the machine-readable form carries the whole comparison in coverage.supersededByRevision at /api/hi/full.json. All six write-ups are published with their method and their failures.
The finding as an image
For a slide, an issue thread or a message.
