How the learning actually works.

A personal dictionary makes an easy promise: dictation that gets better the more you use it. The hard part is the quiet failure that can come with it — a tool that learns something wrong, silently, so you only find out when your text comes out mangled. We wanted learning you could trust, which meant designing it to be visible, patient, and reversible. Here's how it works, without the jargon.

It learns from your corrections, not your commands.

You don't open a settings screen to teach Phrasora a word. You just dictate, and when it mishears a term — a client's name, a framework, the spelling of your own surname — you fix it the way you'd fix any typo. Phrasora watches that correction land and treats it as a lesson: this is the word I heard, and this is the word you actually meant.

That's the natural path. There's still a manual dictionary if you want to add a term up front, but the everyday way it gets smarter is simply you, writing and correcting as usual.

It waits for repetition on purpose.

One correction isn't proof. People misspeak, autocorrect interferes, a one-off fix might not mean anything. So Phrasora doesn't enforce a new term the first time it sees a correction — it waits until the same correction shows up a few times before it starts applying it automatically.

This is the difference between learning and overreacting. A tool that locks in after a single fix feels responsive for a day and then starts inserting words you never wanted. Waiting for a pattern is what keeps the dictionary trustworthy as it grows.

It tells you what it learned.

The moment a term graduates into your dictionary, Phrasora says so — a small chip naming exactly what was added, with a few seconds to undo. The undo is real: it removes the term everywhere, on every device you're signed in on, and blocks it from being immediately re-learned.

That visibility is the whole point. Silent learning is how dictionaries rot. When you can see each addition the instant it happens, a wrong one is a two-second fix instead of a mystery you trace back later.

It remembers what gets confused with what.

A flat list of words degrades as it grows — at a few hundred entries, terms start colliding and accuracy slips. Phrasora's dictionary tracks not just the right word but what each term tends to be misheard as. That context is what lets it stay sharp at scale: it knows that a particular sound, in your speech, usually means a specific name, and applies that only where the evidence supports it.

It's also careful in the other direction. On text where a learned term never appears, it leaves things alone — in our proof test, a dictionary that fixed all fifteen specialized terms made zero false insertions on sentences that didn't contain them.

Some learning never leaves your device.

There's a second kind of learning on the iPhone keyboard, separate from your dictionary: it learns the words and phrases you type to offer better suggestions, and keeps a small sense of which words tend to follow which. That typing memory lives only on your phone — it is never synced to our servers and is wiped the instant you sign out. Your synced dictionary, the one that follows you across devices, is the part tied to your account, and it goes away entirely when you delete that account.

The shape of it.

Learn from real corrections. Wait for a pattern before acting. Announce every addition and make it reversible. Track confusions so accuracy holds as the list grows. Keep on-device learning on the device. The result is a tool that quietly fits your vocabulary without ever surprising you — which is what learning should have meant all along.

Phrasora is free for 2,000 words a week, no card required. See where it runs, or read how your data is handled before you start.