Scribla: Dictation Keyboard

Now on iPad too: beside your mail, in full landscape. Math writes numbers the way your language writes them, and AI takes follow-ups in Spanish and Chinese.

Dictation hands back a wall of text without a single comma. By the time you have put the punctuation in and repaired the names, typing would have been faster. That is why people give up on it.

Scribla lives inside the keyboard. Press record, say it, and the text lands in the field you are already standing in: a messenger, mail, notes, a work chat. Nothing to switch to, nothing to copy back.

PUNCTUATION AND PARAGRAPHS, ALREADY IN PLACE

One long run of speech comes back split into paragraphs, with agreement repaired and stumbles taken out. Fillers that carry weight in written speech are left where they are, on purpose.

FOUR LANGUAGES, AND YOU PICK A PAIR

English, Spanish, Russian, Chinese. The pair is not a list of favourites, it is how the app works: the language key moves between your two, translation runs from one into the other, and the "you spoke the other language" prompt compares exactly that pair. Each language model is downloaded once and then runs with no network at all.

THE NAMES THAT BREAK EVERY TIME

"En dee ay" instead of NDA. A colleague's surname spelled out one syllable at a time. Write the replacement once and it goes in from then on. The keyboard carries its own letter layout for this, because what needs fixing is usually a word you cannot dictate.

EDITING THAT DOES NOT EAT THE GAIN

The edit strip breaks what you said into words: tap one and change its form, its case or its spelling without leaving the keyboard. The space bar doubles as a trackpad: slide sideways and the cursor walks by letters, slide up or down and it walks whole words at a time. One key removes the sentence the cursor sits in, for when the phrase came back wrong. Undo returns the entire insert rather than one character at a time.

SPEED

Speech runs at 120–150 words a minute against 30–40 when typing. And the person at the other end gets text they can read in ten seconds instead of a voice message they have to sit through.

AUDIO NEVER LEAVES THE PHONE

Speech is recognized on the device itself. The audio is not sent anywhere, and not because we promise it: there is no network in the recognition path at all. The keyboard never goes online and never reads what you type on the ordinary keys.

FOUR MODES, ONE KEY

Text. Ordinary dictation with the punctuation put in.

Math. Say "fifteen percent of two thousand four hundred" and get "15% of 2,400 = 360". Currencies convert at the published rate for the day.

Translate. Between the two languages of your pair, both ways. Need one line in a fourth language — hold the Translate key and pick the target.

AI. Ask a question out loud, get an answer as text rather than a transcript of your own question. It answers in the language you spoke.

WHAT GOES OUT

Audio, never. Text, in the AI modes: polishing, answers and synonyms are worked out by a model on the other side, and none of it runs without a connection. Those modes are on from first launch, and one switch in settings turns them off.

WHO THIS IS FOR

People who write a lot from a phone and cannot, or would rather not, send their voice to somebody else's server: lawyers, doctors, anyone working under an NDA. And anyone simply irritated by how slow typing is.