How We Made baba's Hebrew Translations Feel Instant

Why did a translation take two seconds?

For most of this year, tapping Translate in baba took about two seconds before the Hebrew finished arriving. We assumed our own code was the problem, so we measured a full week of real translations before changing anything.

It wasn't our code. The first word arrived about 1.5 seconds after the request reached our server, and roughly 1.2 seconds of that was the language model producing its first word. Our own work before the model, checking who you are and saving the translation, took about three tenths of a second.

When almost all of the wait is the model thinking, there is only one way to make it shorter: start earlier.

What does baba do while you are still typing?

baba now begins translating before you tap. Two moments trigger it:

• When you paste. Pasted text is finished by definition, so the translation starts immediately.
• When you stop typing. After a short pause, baba checks whether your message looks complete, then starts.

The work happens quietly on our server. Nothing appears on screen, nothing is saved to your history, and nothing counts against your usage. When you tap Translate, the finished translation is already waiting, and the whole sentence appears at once instead of word by word.

If you edit the text after that, the earlier attempt is thrown away and baba translates what you actually wrote. If you tap before the early attempt is done, baba runs a normal translation at the same time and shows whichever produces text first.

What you do	What baba does	What you see
Paste a message	Starts translating right away	The full translation on tap
Type, then pause	Checks if you're done, then starts	The full translation on tap
Type half a sentence and stop	Waits, because you are mid-thought	Nothing until you finish
Edit after pasting	Drops the early attempt	A fresh translation of your edit

How does baba know you finished typing?

This is the interesting part. A pause is not the same as a finished message. People stop mid-sentence to think, to switch apps, or to find a word. Starting a translation every time someone hesitates would waste a lot of work and still leave them waiting.

We use TypeSafe Jev for this. It answers one question about your text and returns a probability rather than a guess: is this message finished, or is the person still writing it? Below a threshold, baba waits.

The simple rules we tried first were not good enough. Punctuation does not settle it, because "I'll be there in ten minutes" usually has none. Length does not settle it either, because a single word is often a finished lookup. On a test set of complete and half-typed messages, the model got 40 of 42 right, and both misses were complete sentences it judged unfinished, which only costs a head start.

The check runs while you are still on the screen, so it never delays anything you see.

Who are you speaking to? We stopped asking so often

Hebrew changes with the person you are speaking to. "Are you coming?" is different for a man, a woman, or a group. baba has always let you pick, and after every English to Hebrew translation it asked whether you were speaking to someone specific.

The trouble is that most translations don't change at all with the listener. "Where is the train station?" is the same sentence no matter who you ask.

So we tested it. We took 150 real translations, had them labelled independently, and checked how often the Hebrew actually depends on the listener. The answer was about a third. The prompt was appearing on all of them.

Now baba asks the same model whether the wording depends on the listener, and only shows the question when it does. On our test set it agrees with the human labels about 95% of the time, and the audience control is always there if you want to change it yourself.

What we measured

These are server-side numbers, measured on our production system:

	Before	After
First word of a translation	about 1.5 seconds	about 0.3 seconds when it was started early
Whole translation finished	about 2 seconds	under half a second when it was started early
"Speaking to someone specific?"	after every Hebrew translation	on about a third of them

A translation you tap without any warning still takes as long as the model needs. What changed is that most translations no longer start from zero.

What we threw away

The first thing we built with this was a badge that labelled the writing style of your text: casual, work, formal, marketing, and so on. It worked. On 260 real texts it was accurate about 91% of the time, and far better than the pattern matching we had before.

We deleted it.

Before shipping we tested whether telling the translator about the style made the Hebrew better. It did not. In a blind test of hundreds of translations, the existing prompt matched the right tone every time, and the style-guided version was slightly worse, with a few translations coming out stiffer than they should. The model was already reading the tone from your text, so the label added nothing.

We also tested two other ideas and left them out: offering reply drafts only for real messages, and flagging slang. Both were accurate enough to be tempting and not useful enough to earn a place on the screen.

A feature that measures well and changes nothing is still clutter. The one check that survived is the one that decides something: whether to start translating before you tap.

When can you use it?

The faster translations are live on Android now, and the iOS update is with Apple for review. You don't have to turn anything on. Paste something, or type a message and pause for a moment, and the translation should be waiting when you tap.