ANVE: the private AI office, on the record. See what that means

ANVE

Private AI glossary · updated 2026-10-07

Transliteration
Changing the script, not the language, and why Indian text needs it.

Short answer. Transliteration converts text from one writing system to another while keeping the sound, not the meaning: कल becomes kal, whereas translation would give tomorrow or yesterday. It matters in India because people use many scripts and often type in Roman letters: the Aksharantar project built 26 million transliteration pairs for 21 Indic languages across 12 scripts. For AI it shows up in three places: typing Hindi in Roman letters, search that must match both spellings, and transcripts that must decide whether a word appears as meeting or मीटिंग.

01

Transliteration, translation and transcription

TermWhat changesExample
TransliterationThe script; the sound is keptकल becomes kal
TranslationThe language; the meaning is keptकल becomes tomorrow or yesterday, depending on the sentence
TranscriptionSpeech becomes written wordsAudio of someone saying kal becomes the text kal or कल
02

Scripts in India

The Aksharantar paper lists three kinds of script in use: the Brahmi family of abugida scripts for most major Indic languages, Arabic-derived scripts for some, and the Roman alphabet for many languages with a recent literary history. Several scripts serve more than one language.

ScriptLanguages written in it (examples)Note
DevanagariHindi, Marathi, Konkani, Maithili, SanskritThe Unicode block runs from U+0900 to U+097F
Bengali scriptBengali, Assamese, SantaliOne script, several languages
Arabic-derived scriptsUrdu, Kashmiri, SindhiWritten right to left
RomanRomanised Hindi and other languages typed on phonesMost people are comfortable with the Roman keyboard, so romanised input is popular
03

Why Roman spellings vary

Roman Hindi has no single standard. The PHINC paper gives mujhe, mjhe and mujhee as variants of the Hindi counterpart of the English word “me”, and says this variation makes machine translation harder.

AI4Bharat’s Aksharantar dataset has 26 million transliteration pairs for 21 Indic languages, and a test set of 103 thousand word pairs across 19 languages that separates native words, foreign words, frequent words and rare words. Its IndicXlit model improves accuracy by 15% on the Dakshina test set. The models and data are released as open source.

04

Where it shows up in office AI

  • Typing and dictation: people type Hindi in Roman letters and expect the software to cope.
  • Search: a query for kal should find कल, and the reverse, so the index must handle both scripts.
  • Names: one person’s name spelt differently on a passport, a GST certificate and a bank form is a transliteration problem.
  • Transcripts: in the MUCS 2021 data the same English word appeared in Latin and in native script, so scoring had to accept either.
  • Documents: a Hindi form with English field labels needs both scripts read correctly.
05

What to check in a tool

  • Which direction it converts: Roman to native script, native to Roman, or both.
  • How it handles names and English words inside Hindi text.
  • Whether spelling is consistent from one document to the next; keep a glossary of preferred spellings.
  • Whether search matches both scripts, and how it treats common variants.
  • For ANVE, which script a transcript uses and whether it can be switched is a point to get in writing before a paid pilot.
06

Questions people ask

Is transliteration the same as translation?
No. Transliteration changes the script and keeps the sound. Translation changes the language and keeps the meaning.
What is romanised Hindi?
Hindi written in Roman letters, as people type it on phones. Mixed with English words it is what people call Hinglish.
Can AI convert Roman Hindi to Devanagari?
Open models built for this exist: AI4Bharat’s IndicXlit, trained on Aksharantar’s 26 million pairs, covers 21 Indic languages. Accuracy varies by word type, so review names and rare words.
Why do search results differ between kal and कल?
Search matches the characters it is given. A system that indexes only one script will miss the other, so check that it handles both.
07

What is not real yet

No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.

Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.

08

Sources