What DictaBooth keeps on your device, what it sends, only when you press something, and what it never does. No account is needed, and nothing here measures you.
Last updated 11 November 2026
Unless you sign in to connect devices: then a copy of your progress is kept for you (signing in).
Each is used to answer you; what you wrote or said is not written to disk or logged.
What you type for dictation is never sent anywhere. The marking, the blanks and the score are all worked out in this browser, so the one thing you do most of here leaves no trace off this device. Writing that is meant to be checked (a Paraphrase version, a talk, a question to Dicta) is sent only when you send it.
What you write in Paraphrase, and the colours it earned, are kept in this browser against that video, like your report, and are not uploaded without an account. Forget this video removes them with the rest of that video's record.
Unless you sign in (below), transcripts you fetch or paste, where you reached, your saved words and your settings are kept in this browser only. They are not uploaded, and not shared with anyone else using this page. Forget this video on the first screen removes one; clearing site data removes the lot.
Your report is part of that too, and is the most personal of it: the words you got wrong, what you wrote instead, and how long the speaker gave them. It is kept with your place in that video, in this browser, so the report is still there the next time you open it and can say which words have caught you twice. Without an account it is never uploaded, and nothing on any server knows that you found a word hard. Forget this video removes it with the rest of that video's record; Again on the report starts the video, and its report, from nothing.
The words you save are part of that. Each one holds the word, the meaning the dictionary gave, the sentence you met it in, which video and which second that was, the language, and the dates deciding when it comes back to you. Those dates are worked out here, in this browser, from your own answers on the cards and what you say in talks, so no server knows which words you find hard, or that you are reviewing at all. Removing a word removes its record; clearing site data removes them all, and without an account there is no copy anywhere else to remove. Each word also holds its place on your memory path: its level from 1 to 6 and when it is due, worked out here in the same way.
Exporting them is your browser writing a file and handing it to you. The flashcard deck, the Anki notes file, the Notion CSV, the plain CSV and the Markdown are all built here and never sent anywhere. Once a file is yours it is outside this policy: it holds the words you saved, so where it goes next is your decision rather than ours.
The flashcard deck is a page of its own, and it
asks nothing of anyone when you open it: no fonts, no code from elsewhere,
no pictures from this site, nothing at all. The one exception is pressing
play on a card, which asks YouTube for that moment of that video, from
youtube-nocookie.com, the same host the player here uses.
Before you press it, nothing is asked. The deck also keeps its own review
dates in whatever browser you open it in, separate from this site's, and
the page says so on itself.
Your practice journal is a running record of every try: what was asked, what you answered, whether it was right, and when, with how it went (the playback speed, how often you replayed a line, a spoken take's scores), and the words you look up while reading or listening. It is how your history survives what the report keeps (the report holds only the last mistakes of each video). It is kept in this browser's database, never uploaded, and is not part of connecting devices. From it this device also works out which grammar points each try was evidence of, and from that your level and your plan in the report; that stays here too, and is rebuilt from the journal, never sent. So do your goal, your choices in the plan (and the reason, if you give one), the record of what the plan proposed and what came of it, the result of the 5-minute check and of each check-up: kept in this browser only. Download it, in the Save a phrase panel, hands you a copy as a file; clearing site data removes it.
Asking Dicta. The box on the first screen answers the common questions (what you can do here, what to learn now, which videos, your progress) from what this browser already holds, and sends nothing. Any other question goes to our helper with a short summary the page builds from your practice: the language, the skill levels measured so far and the weakest of them, how many words you have saved and how many sit at each level, up to five new words with their meanings, how many cards are due, your goal and the titles of your last three videos, with your question and the last few turns of the conversation. The helper passes it to Groq for an answer. It does not write it to disk or log it; it counts the questions per day and nothing more, and the conversation itself stays in this page.
When a service cannot answer. If Groq or OpenAI cannot answer a request described on this page, our helper may send the same request to Anthropic (its model Claude) instead, under the same rules: it is not written to disk or logged, and it is used to answer and dropped.
Paraphrase works in two steps, and sends something at each. Opening a stretch of a video sends that stretch's own text, which is the video's and not yours, to our helper, which asks Jev (TypeSafe's fast yes/no model) which words matter and Groq (a language-model host) to write the note of how the stretch is built. That note is kept on the helper against the text, so the next person to open the same stretch costs nothing, and it holds nothing about anyone. Pressing Check sends your version, with the video's text, to the helper, which asks Anthropic (Claude) whether you kept the ideas and what to fix; when Claude cannot answer, it asks Jev whether you kept the ideas and where the grammar is doubtful, and sends the sentences Jev doubted to Groq to find the exact wrong words. The helper does not write your version to disk, does not log it, and does not put it in its problem log; it is used to answer and dropped. Anthropic, Jev and Groq receive it to answer the request, and what either does with it afterwards is governed by its own policy, not this one. The helper counts the number of such requests per day and no more.
Talking with Dicta. Each thing you say or type in a talk is sent to our helper with the talk so far, its topic and your level, and the helper asks Anthropic (its model Claude) for Dicta's reply and the feedback at the end, and Groq for the shorter help (a topic, a hint when you ask for one), or when Claude cannot answer. A spoken turn is recorded only while the microphone is on and is turned into text by Groq first; it can also be scored for pronunciation, as in shadowing. The helper does not write the talk to disk or log it; the talk and its record stay in this browser.
The report's exercises (German for now) work the same way. Reading questions on a story are made by our helper with OpenAI from the story's own text and kept against it; your answers to them never leave the browser. Write about it, Answer aloud and Add your own work send your answer (a spoken one first as a recording, turned into text by Groq; a photo of your homework or workbook first read into text by Claude, Anthropic's model, or Groq's vision model when Claude cannot) to the helper, which asks Groq to score it and find the wrong words. You see the text read from a photo and correct it before anything is checked. The helper does not write your answer, recording or photo to disk and does not log them; they are used to answer and dropped. A photo is made smaller on your device before it is sent.
Your saved words and talks. To group your saved words into topics, the page sends them (each with its meaning and example sentence) to our helper, which asks Anthropic (Claude) which situations each belongs to; to put an idea on each topic on the first screen, it sends the topic names and their words; and a speaking simulation sends the words it is planned around. To suggest what to talk about, it sends your level, the titles of the videos you practised in that language and some of your saved words, which the helper passes to OpenAI. The helper does not write them to disk and does not log them; they are used to answer and dropped. The topics and ideas are kept in this browser.
A dictionary is asked only when you click a word in the answer key, and is sent that one word. With the helper server running it does the asking; without one your browser asks Wiktionary directly, so your IP address reaches them the way it would if you opened the page yourself. For Chinese, the sentence is also sent to the helper so a click on one character can be widened to the whole word.
Adding a word by hand works the same way. When you press Find examples, the word you typed and its language are sent to the helper, which looks up its meaning and asks Tatoeba, an open sentence collection, for sentences that use it. Nothing else from the form is sent: the meaning and sentence you keep are saved in this browser only.
Saving a phrase sends the phrase, the sentence you met it in (if you
gave one) and its language to the helper once, which asks OpenAI what the
phrase means in that sentence. Not where you met it, not the page's
address, nothing else. The helper keeps the answer in memory for a while so
the same phrase in the same sentence is not asked twice, and never writes it
to disk or to a log, because a phrase may come from a private chat. The
phrase itself, where you met it, and when it comes back to you are kept
with your saved words: in this browser, and in your connected copy only if
you choose to connect devices (below). The Save to DictaBooth bookmark only opens this
page with the phrase in the address after the #, which browsers
never send to a server. The same goes for Paste a phrase (your
clipboard is read only when you press it, and only into the form on this
page), for Share on an Android phone where DictaBooth was added to
the home screen (the phone turns what you shared into that link itself,
before anything is sent), and for the Save to DictaBooth iPhone
shortcut, which opens the same link.
Your microphone is used only while you press for it, and never otherwise. Where the recording goes depends on what you pressed:
The tone check on a Chinese word stays here: the sound becomes a pitch line inside the browser and is then dropped. Nothing is uploaded and no recognition is involved.
Shadowing, saying a whole line back to be scored, does upload. The recording is sent to our helper and on to Microsoft Azure's pronunciation assessment, which returns the scores you see. It is held in memory for as long as that takes, is never written to disk, and is not kept afterwards by us. If you would rather not have your voice leave this device, do not press it; nothing else on the page uses it.
Check my pronunciation, saying anything you like to be scored, works like shadowing: the recording goes to our helper and on to Microsoft Azure, which scores it with no line to compare against and sends back what it heard. It is not written to disk or kept. A word that came out weak is saved to your words, in this browser.
Saying a word on a flashcard is scored the same way, and only the score is used, to decide the card.
When the day's allowance is gone, shadowing does not stop. Your browser's own speech recognizer checks the words instead, and that is not more private, only differently so: Chrome and Edge send the recording to Google to be recognized. A different company rather than no company, and only on a press you make knowing this. Firefox has no such recognizer, so there nothing is sent and the shadowing waits for tomorrow.
In a talk, and with the mic in the box on the first screen, what you say is turned into text first; both are described under Talking with Dicta and your own texts.
Speaking your Paraphrase version is optional. Tapping the microphone records what you say, only while the listening bar is on screen (a minute at most), and sends that recording to the helper, which passes it to Groq (the same host, running the Whisper speech model) to be turned into text. The text comes back into the box for you to correct, and nothing is checked until you press Check. A recording with no voice in it is not sent. The helper does not write the recording to disk and does not log it or the words; it is used to answer and dropped. Groq's own policy governs what it does with the audio afterwards. The helper counts the length of such recordings per day, and when the day's allowance is used up the microphone waits until tomorrow and typing carries on. Your phone's keyboard also has its own microphone key, which is the phone's own dictation and has nothing to do with this page.
Your own text. A text you paste, or a PDF or Word file you add, is read in this browser, and it is kept on this device. Once you open it, the part you are reading is sent to our helper to be translated, which passes it to OpenAI and keeps nothing; the translation is kept on this device. A word you tap in it, or in a talk, can go the same way, to Groq, for its meaning in the language you read translations in. What you add to start a talk, up to its first 1,500 characters, is sent to our helper, which passes it to Groq or OpenAI to plan the talk and keeps nothing. A photo, pasted in or added, is sent to our helper to be read by Claude (Anthropic), or Groq when Claude cannot, and is not kept. An audio file you add, or what you say into the mic in the box, goes to our helper, which passes it to Groq or OpenAI to be written out and keeps nothing. A web link sends the address to our helper, which fetches that page's text once for you and keeps nothing.
YouTube plays the video, so it sees that you asked for it, and it
may set cookies. Nothing is requested from YouTube, not even its player
code, until you allow it. The player is loaded from
youtube-nocookie.com where YouTube accepts it, which sets
fewer cookies than the ordinary embed.
Fetching a transcript sends the video's id to our helper server, which asks YouTube for the captions. Transcripts are cached there so the next person to open the same video costs nothing to serve. A cached transcript is a public video's captions, and it holds nothing about you, and it is not tied to who asked for it.
Stories to practice on asks our helper for popular videos in the language you chose, and for a video's transcript when you press Read. The pictures on the news cards are kept by the helper too, so your browser does not contact YouTube for them, and the last list you saw is kept in this browser so it appears at once. Only the language and the video are sent, nothing about you. The Easy and Medium levels send that public video's transcript to an AI service to be rewritten for reading; the rewrite is stored on the helper for the next reader, and holds nothing about you. Hard plays the video itself in YouTube's own player, and only once you have allowed YouTube. Easy and Medium are read aloud by your own device's voice, and nothing is sent for that.
Up next, the videos suggested while you watch, comes partly from this browser (the videos you left half-way) and partly from our helper: it is sent the id of the video you are watching and finds that creator's other videos (only that listing reaches YouTube, and it is kept for days), and the stories come from the same list as above. Your level is matched here, in this browser; nothing about you is sent.
Transcribing the audio, which happens only if you press the button offered when a video has no captions, downloads that video's audio to the helper and sends it to a speech-recognition service to be turned into text. The audio is the video's, not yours; it is deleted once nothing still needs it. Transcribe on this device is the alternative: it downloads a recognition model to your browser and the audio is never sent anywhere.
Reading subtitles off the screen, offered when a video has no captions but prints them on the picture, sends cropped frames of that video to be read by Claude (Anthropic), or by Groq's vision model when Claude cannot. Again the pictures are the video's, not yours.
Describe a picture sends the talk's topic, its language and level, and which pictures you have already had to our helper, which searches Openverse and Pexels, two free photo libraries, for a photo. The photo is then loaded from them, so they see your IP address the way any site does.
Joining is your choice. Your report may ask learners of German to help test whether talks make new words stick. Nothing happens unless you press Join. For 35 days, the new words you save are split by a fixed rule into two groups: half are planned into talks as usual, half only come back as flashcards. You can still use any word. The split comes from your study record, which travels with your account like your saved words, so it is the same on every device you sign in on.
What is sent: once a week, the page sends our helper counts under a random study id: the language, your talk level, and for each group how many words were saved, how many are at each level, how many reached level 5 and in how many days, plus how many talks you did and minutes you practised that week. No words, no sentences, no recordings and no video titles, and nothing tied to your account or e-mail address.
What is kept: one small file per study id on our helper. It is not logged, and your IP address is not kept with it (the helper counts requests per address for the day, in memory, as it does everywhere). Leave in your report deletes the file; the page asks again until the helper confirms. Files nobody has sent to for 120 days are deleted too. When the check ends, only the totals across everyone are published.
Two other places your browser reaches, both of them for files rather than about you: Google Fonts serves the typeface, and jsDelivr serves the on-device recognition model if you ever ask for one, and the small readers the page needs when you add a PDF or an Anki deck (the file itself is read in your browser and goes nowhere). Each sees your IP address the way any site's host does. YouTube's thumbnails on the "carry on" cards wait for the same permission the player does, so before you give it nothing is asked of YouTube at all.
Limits are counted per IP address so one visitor cannot spend a day's transcription on everyone else's behalf. The count lives in the helper's memory, resets daily, and is not written down or joined to anything else.
You never have to. Without an account, everything above holds and your progress stays in your browser.
Creating an account keeps your username, your e-mail address, your password (only as a salted scrypt hash: nobody can read it back), the date of the Terms you agreed to, and the devices signed in. We e-mail you from support@dictabooth.com, through the mail service that already handles our domain's e-mail (Hostinger), only to confirm your address or reset your password: no newsletters, and the address is not shared.
Continue with Google is there only if you press it. Google's sign-in script is loaded only when the sign-in box is open. Google tells us your e-mail address and that it is checked, and Google learns that you signed in to DictaBooth. We do not get your Google password or contacts.
While you are signed in, the helper also keeps a copy of your progress, so your other devices can read it. That copy holds your places in videos with what you typed and said in them and the words you missed, your saved words with their reminders and levels, the words you looked up with their meaning and a short sentence they were in, your Paraphrase answers and fixes, and your points and days practiced, and your plan: its goal, your choices, its record of results, your check-ups and your 5-minute check. It does not hold your practice journal (it is too large, and stays on that device), the transcripts, video audio or your settings. Nobody reads it, it is not used for anything else, and it is not shared. A link we e-mail works once, for 15 minutes. Sign out ends it on that device. Delete account removes your account and the progress kept on the server (Delete synced progress, as it used to be called) and signs out every device, and leaves what is in your browsers alone.
There are no analytics, no advertising and no tracking pixels, so there is nothing here that builds a picture of you over time.
DictaBooth needs no account, has no analytics and no advertising, so there is no profile of you to hold, sell or lose. If you choose to make an account, your account details and the synced progress described above are kept for you, and you can delete it. The things that do leave your browser leave because you pressed something that needed them to, and each one is named above.
This is a small free tool, not a company. If something here is unclear or you think it is wrong, the fastest fix is to say so, and the page will be corrected.
Write to support@dictabooth.com. If you email us for a transcript, the teachers and creators page says what we do with your address: answer you, and nothing else.