← appaday

It hears what you actually said

My Hungarian is, in my own words at the time, a broken beginner to intermediate. What I could not find was anything to simply talk at — out loud, badly, for ten minutes — that would talk back and not turn every sentence into a lesson. So I asked for exactly that: the idea is that I can freely talk to an AI voice.

It is called Beszélgetés, which means conversation. You hold the space bar, say something in Hungarian, and let go. Anna answers out loud.

A dark page headed Talk Hungarian to someone who talks back. An example turn: the learner says Tegnap mege a boltba és vettem... strawberry, meaning yesterday I went to the shop and bought strawberry. Anna replies Epret vettél? Én is imádom az epret. Mit csináltál még tegnap, meaning you bought strawberries, I love strawberries too, what else did you do yesterday. Below, two notes: mege corrected to mentem, past tense of megy, first person; and strawberry in English mapped to eper, the word you reached for, saved to your list
One turn. She never says that was wrong — she just happens to use mentem and epret in her reply.

The one decision everything else follows from

Almost every speaking app works the same way: transcribe what you said, then grade the transcript. It sounds reasonable and it is quietly fatal. A transcriber is trained to produce correct Hungarian, so it repairs you on the way in. You say mege a bolt, it writes mentem a boltba, and the mistake you needed to see is gone before anything gets a chance to look at it. Pronunciation feedback is impossible for the same reason — a transcript has no accent in it.

So here the recording is never transcribed. It goes straight into a model that listens to audio, and comes back as what you actually said: errors intact, and the English or German word you grabbed when you got stuck left exactly where you put it.

The reply arrives as text after about eight seconds and as a voice roughly five seconds later, on purpose — synthesis takes about as long as thinking, and there is no reason to stare at a blank screen for both.

Corrections stay out of the conversation

Anna is instructed never to announce a correction, only to recast: use the right form herself, naturally, in whatever she says next. The real teaching piles up in a side panel and in a short review when you stop, so the talking stays talking. Click any correction to hear it said properly.

The words I reached for

The idea I like most came later: a list of every word I had to say in English or German because I did not have the Hungarian — sorted by how common that word actually is.

The ranking is corpus data, not a model's opinion: the top fifty thousand words from film subtitles, where line number is rank. Hungarian stacks suffixes onto everything, so each captured word is first reduced to its dictionary form plus a few common inflections and the best-ranked one wins. That is how érez, to feel, lands at #309, while kormányablak — the government office you queue at — sits somewhere past fifty thousand. Learn from the top down.

It costs money, so it has a door

Every turn costs real money, which is fine until a stranger finds it. So there is no account, but there is a code: enter it once and the browser remembers. Each code has its own daily allowance and its own private history, and the site as a whole has a ceiling so that one busy day cannot break it for everyone. History lives on the server, so a conversation started on the phone carries on at the desk.

A list headed Honest limits: Anna is a language model, not a teacher, she mishears sometimes and a correction can be wrong or fussier than a native speaker would be. She takes a few seconds, reply in text after about eight seconds and voice a few seconds later. Pronunciation notes are only as good as your microphone and she is told to stay quiet rather than guess. Tested in desktop Chrome and Edge, phones not yet the tested path. Recordings are sent to Google's Gemini API to be understood and answered and are not kept afterwards. The daily budget can run out mid-conversation. One person built this and runs it in his spare time
The landing page's own list of what is wrong with it.

Still rough

She mishears. She is occasionally fussier than any real Hungarian would be. It is a conversation with a slightly slow friend, not a phone call, and it has only really been tested on a desktop. But it stays a conversation, which is the part every other tool I tried gave up first.

There is a free demo code on the page, good for a short conversation a day for a week, with no sign-up.