Your decks, your settings and the audio the dictation tab generates all stay on this device. Nothing is uploaded and there is no account. Saving starts immediately, in this browser's own storage — you need not set anything up. It is kept across reloads, but only for as long as this page's site data lasts.

Chrome and Edge can do better: pick a folder and the same files sit on disk where you can see them, copy them, and put them inside Drive or Dropbox for sync. Choosing a folder moves everything already saved into it; disconnecting moves it back. Firefox and Safari have no folder picker, so browser storage is the whole story there.

One file holds a whole setup: every deck and all your settings, ready to keep somewhere safe, carry to another machine, or bring back after a browser clears its storage. The API key is left out on purpose, and so is the dictation audio — it is a cache the Dictation tab refills, not something to mail around.

A backup is the whole store, audio and all: a zip laid out exactly like the data folder. It is the one copy that keeps the dictation bank, which cost real API calls to make. Unzip it anywhere and Chrome can open it as a data folder; or bring it straight back with Restore, which overwrites any file of the same name.

Dictation, Shadowing, Reading, Writing, Translate and Conversation call the Gemini API directly from this page with your key. It is stored in this browser's local storage, never alongside your decks, so it cannot end up in a git commit, and it is sent to nobody but Google. Get one free at aistudio.google.com/apikey.

The Azure Speech key, for Azure's read-aloud voices, is set under Read-aloud voice, where it is used. It is kept in this browser only too.

Models

Every Gemini model you use, listed once each. The limits belong to the model, not to the job it is doing — Google counts calls per model, so a model given two jobs below draws both of them from the one allowance you enter here, and typing it twice would only be a way of lying to yourself about what is left.

Calls / min and calls / day are counted in this browser before a request goes out, so you never discover a limit by being refused by Google. The windows roll over the last minute and the last 24 hours rather than resetting at midnight, so the day's budget trickles back an hour at a time. 0 means unlimited — the right answer for a paid key, and the one case where Google gets to be the one who says no. The defaults are the free tier.

Who does what

Each job's model is chosen in the section of the mode it belongs to: Dictation's in Dictation, Conversation's six in Conversation, and so on. The list above says which jobs each model is doing.

Pick the same model for two jobs and they share its allowance, exactly as they already do at Google's end. The two speech jobs are the ones that really have to differ: only the TTS models return audio. Conversation replies suit a flash-lite model with a bigger daily allowance: one conversation spends five or six of its calls, each a line or two.

If a model ID is wrong the API's own error is shown verbatim, so a 404 here means exactly that. There is deliberately no hidden fallback list: the model that answers is the one you typed, and nothing quietly spends on another.

Call budget

What has actually been spent, against the limits set under Models. One line per model, not per job, because that is how the counting works: two jobs on one model run one counter down together.

The first note shapes the words the text model writes. The audio note is added to every request that makes audio with a Gemini voice, the dictation sentences and a Reading text read aloud, as a line of direction before the text. Your device's own voice and Azure voices do not use it.

Any language the model writes is fine; type its English name. Answer checking is language-agnostic: it compares accents and diacritics exactly, and aligns words with them stripped, which works the same for Vietnamese tones, Spanish acutes and German umlauts.

How every note and every piece of feedback should be written, in your own words: Shadowing's notes, and the feedback in Writing, Translate and Conversation. A language (English), several (English and Vietnamese), or more than that (English, avoid technical terms, Vietnamese, with an English summary at the end). The words you are practising are always quoted as written. Left empty, the notes are in English.

What reads words aloud in Typing, and the flashcard lines in Shadowing. Neither choice is Gemini, and neither spends anything from your Gemini budget. Dictation and Reading have Gemini voices of their own, in their sections.

Select a word or a short phrase anywhere on the page, such as a new word in a dictation sentence, and a small form opens under it. It shows the word, its translation and an empty notes box, and Add puts the word into the deck below as a new card. If that deck already has the word, the form shows the card as it is stored and the button reads Update. Updating changes the meaning, the notes and whether the card is a grammar pattern, and keeps the card's score and history. Tick Grammar pattern for a construction such as hễ … là …; it ticks itself when the word has a … in it. You can select text in either language: text in your own language is taken as the meaning, and the word is looked up from it.

The translation comes from Google Translate, through the free public address its own website uses. It needs no key and spends nothing from your budget. It is also unofficial: Google can refuse it for a while, especially on a busy or shared network, and then the form says so and you type the meaning yourself. The selected text is sent to Google. Ask for notes is Gemini: it spends one call on the Notes model below and uses the notes prompt below it.

The voice that reads the word is set under Read-aloud voice, above: your device's own, or Azure's.

How long a dictation sentence is, and how many of your cards each one is written around.

Gemini voices

One is drawn at random for each new sentence, so listening practice is not always the same speaker. Untick the ones you would rather not hear. You cannot untick them all, since something has to read the sentence.

A set is some lines to read aloud and record. Both sources are drawn from what you already have, and nothing on the Shadowing tab ever writes a new sentence — the only thing that does is the Dictation tab's New sentence button, and both tabs share one bank, so whatever it makes is immediately shadowable.

Each hand-in costs one call, however many lines are in it, because every recording you choose goes up together. You can hand in a whole set at once, or part of it now and the rest later, or just one retake. A set is whatever is actually available up to the number above, and the tab says so when it comes up short. The call is spent on the Shadowing model below, out of that model's own allowance.

What the feedback listens for, one rule per line, for the language above only — every language keeps its own. A line starting style: is about how the notes are written rather than what is heard. The rules are drafted for your language the first time you hand in a set, and each note says which ones it applied. Rate the notes on the Shadowing tab, and once enough notes under one version are rated, with at least one not useful, the rules are rewritten from what you said. Drafting and revising spend one call each on the Listening rules model below. A line you write or change here is yours: revisions keep it word for word.

Shadowing prints the sentence in full, which a dictation deliberately hides. Sets prefer sentences you have already typed in Dictation, and Dictation prefers ones you have not yet shadowed, so the two tabs stay out of each other's way — but a sentence you shadow first is one whose dictation you have already read.

How long an opinion piece should be. A summary asks for the percentage above of that, more for a long text and less for a short one.

A conversation held turn by turn spends one call on the scene model, one on the replies model for each of the model's turns, and one on the feedback model at the end. A spoken turn adds one on the listening model, and a roleplay with spoken turns is graded by the listening model, since it listens to them.

A live conversation, roleplay or find out, spends one call on the scene model, one on the live model for the whole conversation, and one on the live feedback model, which listens to the recording of your microphone. The live model has to be one of Google's Live API models (the audio streams both ways through it as you talk), so it is the one job no other model here can do. Google's free tier limits it by how many live conversations run at once rather than by calls, which is why it starts with no limit here; a number per day entered under Models and budget is a number of conversations a day. The audio goes straight from this page to Google, over your own key, like every other call.

Puts every prompt on this page back to the one this version of the app ships. Your other settings, your listening rules and your decks are not touched.