Tamil Voice Typing

This is the thing that makes Tamil dictation different from dictating Hindi or Russian. Tamil is diglossic: the language people speak (koduntamiḻ) and the language people write (centamiḻ) are separate registers, not an accent apart.

Recognition transcribes what it hears. Say the everyday form and you get the everyday form on the page — கொடுங்க where a letter would want கொடுங்கள், வர்றேன் where writing wants வருகிறேன். For a WhatsApp message that is right. For an essay, an application or anything formal it is a first draft that needs a pass.

The practical trick: dictate in the written register if you can. It is slower to speak that way, and it removes most of the editing.

Tamil’s near-homophone sets are where dictation errs, and they are worth checking deliberately rather than reading past. The three nletters — ண ன ந — and the three l letters — ல ள ழ — are distinguished by tongue position that fast speech blurs. Vowel length (அ against ஆ) goes the same way.

Speaking more slowly genuinely helps. Proofreading those specific letters helps more, because they are where the errors concentrate rather than being spread evenly through the text.

The recogniser is asked for ta-IN — Indian Tamil. Sri Lankan, Malaysian and Singaporean Tamil are understood, since the written language is shared, but region-specific vocabulary is less reliable and will sometimes come back as the nearest Indian-Tamil word instead.

The microphone is the browser’s. On Chrome and Edge the audio goes to the browser vendor’s speech service and text comes back; WorldTyping is never in that path and never receives audio. The text stays in your browser, saved locally so a reload does not lose your work.

That is also why dictation needs an internet connection, and why a browser without Tamil speech support will not respond to the button at all.