Privacy, in plain words

HomeTongue listens to your accent and guesses where you grew up, in any language. Here's exactly what happens to your voice.

By default: your voice and words are never kept

When you talk, your recording is sent to Google's Gemini model, held in memory long enough to answer one question about it, and discarded. It is never written to disk. There is no separate speech-to-text service: the model listens to the audio directly. It does write down what it heard as part of its answer, and that transcript is shown to you on the result card. It stays in memory with everything else and goes when the answer does, unless you choose to donate the clip. No account, no name, no email, no tracking pixels. If you answer "did I get it?", we keep an anonymous scorecard only: the country we guessed, the broad region, the confidence score, whether you said we were right, the country you corrected us to, and a coarse device type (phone or desktop). Never your recording, never your words, never a city. Your own browser also keeps a short list of the answers you gave, so that it stays on your device and not with us; clearing your site data clears it. If your connection drops while we are answering (you put the phone down, or switched apps), that one answer waits on our server for up to thirty minutes so you can collect it when you come back, and it is deleted the moment you do. It is stored encrypted. Your browser generates the key and keeps the only lasting copy of it; it is sent with your request so we can lock the answer and unlock it again when you come back, and it is never written down anywhere on our side. So the thing actually sitting in our database is unreadable on its own, and once you collect your answer the row is deleted.

How we count visits

We measure how the site is used with our own code, not a third-party analytics service. When you press record, get a verdict, pick a deck or finish a game, your browser sends us a one-line note saying so: the name of the action, the page, a coarse device type, a couple of small details such as how many seconds you talked or which deck you chose, the country Vercel's network reports for the request, and, if you arrived from a link, the name of the site it was on (only the site, never the page or address). No cookie is set. Your address is never stored; it goes into a one-way hash with the day and a server secret, so we can count how many people visited today without being able to find any of them. A random session id lives in one browser tab for one visit and is gone when the tab closes. None of these notes ever contain your words, your voice, your nickname, or anything you typed.

If you press "donate this clip"

We additionally keep that one recording, its transcript, your correction including the city if you typed one, and the guess details above, with a timestamp. We use it for one purpose: training a model that recognizes accents better, especially at the city level. Donated clips may become part of a research dataset; they are never sold, never used for ads, and never linked to your identity: we don't know who you are.

Deleting a donated clip

Since there are no accounts, tell us roughly when you donated (date, time, and the guess the app made) and we'll find and delete it: email mohd.e.arab@gmail.com, or open an issue.

Third parties we use

Accent analysis runs on Google (Gemini). Hosting is Vercel. Supabase is the database behind all of it: the anonymous scorecard, the leaderboard, any clip you report, and the short-lived sealed answer described above. Donated clips live in its file storage. Each processes your audio only to provide the service. HomeTongue used to send audio to Groq (Whisper) for transcription and Anthropic (Claude) for analysis; that pipeline was replaced in August 2026 and neither receives your audio any more.

Guess the Voice, the listening game

The game is separate from your voice and never records you. If you post a score, we store the nickname you type, the score, and which clip and distance each round was. That is the leaderboard, and it is public. If you flag a clip as wrong we store which clip, the reason you picked, and any note you type. Pick any nickname you like; we have no way of knowing who you are, and there is no account to delete.

Most decks stream their clips from YouTube so the original creator keeps the view and the credit. Those are real YouTube embeds, which means YouTube can set its own cookies and see that a video was played, exactly as if you had opened it on youtube.com. YouTube only appears on the game page, never on the page where you record your voice. Two others are on every page, including that one: the map tiles come from CARTO and the fonts from Google, so both see that a browser at your IP address loaded a file. Neither is sent your voice, your words or your result.

Last updated August 2026