The story of Armath Input Intelligence
From an idea to a language model of its own.
Every language has roots. Every touch has a pattern. Armath was born where those two worlds meet, and its name says so. Արմատ (armat) is the Armenian word for root: a connection to language, heritage and identity. Math is the mathematics behind algorithms, neural networks and the power of modern computing systems. Together they make Armath: not merely a name, a way of working.
It began with a keyboard that worked
The first weeks went into the things nobody notices until they fail: a key that responds the moment your finger lands, a touch near the edge of a key going to the right letter, the Armenian full stop «։» where it belongs, a preview over each key, an emoji panel with skin tones, a one-handed mode.
Later came the feature that set the direction: a whole sentence in one stroke. You swipe a word, pass over the space bar and keep going. It showed what the project wanted to be: not a keyboard translated into Armenian, but one that thinks about how Armenian is written.
A laboratory before features
Soon the project gained something most keyboards never have: a laboratory of its own. The Algorithm Lab replays thousands of Armenian words, typos and swipe strokes through the keyboard’s real code and reports exactly what changed. From then on an idea was not an improvement until the laboratory said so.
Living inside the limits of iOS
A keyboard on the iPhone is a guest in every other app, and iOS gives it very little memory. We learned that the hard way, and early: loading the language data the ordinary way was enough for iOS to close the keyboard without a word. The answer was optimisation, optimisation and more optimisation.
At the same time the keyboard learned to keep its hands off what it should never see. It does not learn from passwords, one-time codes, card numbers or addresses, and it follows each text field’s own rules about correction and capital letters. And the neural reading of swipes, running entirely on the phone, was measured on a physical iPhone for the first time.
Its own words
Armath left its borrowed vocabulary behind and built one of its own, out of the Armenian language itself. We went through open sources and collections of Armenian web text, counted every word form, kept today’s spelling and kept names as names. The result is a dictionary that speaks living Armenian: the Armenian people write today, in the news, in blogs, on forums and in messages to each other.
That road taught us the most important lesson of all: honesty before numbers. We found that the old list had, without anyone intending it, known the very texts we tested the keyboard on, so it made our results look better than they were. We did not hide it. We wrote it down, rebuilt the dictionary and measured everything again, every word of the test texts, familiar or not. The keyboard now mistakes a correctly typed word for a typo more than four times less often. This is the keyboard Armenian deserves: not one that looks good in a report, but one that holds up to an honest test.
Its own language model
Then Armath gained the piece its manifesto calls its core: a neural language model developed and trained from scratch for Armenian text input. It learned from 531 million words and signs of Armenian web text on a cloud graphics processor, fits in 17.6 megabytes and runs inside the keyboard itself, on the iPhone’s own processor, in about a millisecond per word.
Soon it was given a vote wherever a keyboard has to guess: among the candidates of a swipe, in what to correct at the space bar, in the order of the spelling hints and in every branch of a sentence swipe. Each of those votes went in only after it had been measured to help.
Today we have access to leading solutions in computing systems, and we are working to expand the capacity already present in Armenia, built on NVIDIA processors. More computing power means a more accurate model, and a more living Armenian on the screen of your phone.
What comes next
The more you swipe and share your swipes, the better the learning becomes: not only for you, but for the whole Armenian language. Together we can keep Armenian from falling behind in the digital world. Armath Input Intelligence, a language model that keeps learning, will make Armenian more competitive in today’s world. Help us now.