What happens to your data
ZumiScribe is a web page. It has no server and no account of its own, and it does not keep your audio unless you switch that on. This page says what stays on your computer, what is sent to the speech service you choose, and what you are responsible for checking.
What stays on your computer
- Your API keys. Kept in this browser only, in the way you choose in Settings: until the tab closes, on this device as plain text, or on this device locked with a passphrase. They are sent only to the provider they belong to.
- Your transcripts. Saved, as you go, either into the folder you chose (real files on your computer) or, if you chose none, into this browser's own storage. Nothing is uploaded to ZumiLabs.
- A recovery copy of the latest transcript, and your settings, corrections and notice wording, in this browser's storage.
- Abbreviation look-ups. Words in capitals (such as PXP or API) are underlined in the transcript. Nothing is looked up until you click one. Then only that word, not the transcript, goes to the company glossary address if you set one in Settings, and otherwise (or if that fails) to the public Free Dictionary at freedictionaryapi.com. The same goes for any word or phrase you right-click and choose Look up: only that text, and only then, goes to the company glossary, then the Free Dictionary and Wikipedia (en.wikipedia.org). Its "Search the web" link opens DuckDuckGo with the text only if you click it. The company glossary and the public sources can each be switched off in Settings, under Glossary.
- Spell check. Off unless turned on in Settings, Auto Correct. It is the browser's own: its basic spell check stays on this computer, while Chrome's "Enhanced spell check" (a Chrome setting) sends the text being checked to Google.
- Sharing captions. Only if you switch it on (Settings, Sharing): the caption text, speaker names and title are sent to the follow-along relay you run, and from there to every phone that opens the link. With a tunnel (Cloudflare or ngrok) the text passes through that service. Anyone with the link can read the captions as they are spoken, so treat the link like a meeting invitation and tell people it is happening. A PIN can be required. Audio is never sent.
- Translation. Done by the browser's own translator on this computer (Chrome 138 or later): a language is downloaded once, then the words are not sent anywhere. The translations are saved in the recording's file next to the original words.
- The capture area. Which part of a shared screen holds the slides is kept in this browser and in the recording's file. It is only a rectangle; the screen's picture is never sent anywhere by it.
- Summaries. Chrome's built-in AI and Ollama keep the text on this computer. If you press the summary button, the transcript text (with the title and attendee names) is sent to the model server named in Settings, Summary. By default that is Ollama on this computer (127.0.0.1), so it stays on this machine; if you point it at another address, the text goes there. Unless you choose OpenAI or Gemini as the summary writer (Settings, Summary): then pressing the button in the summary dialog sends the same text to that company, with the key you saved for it, and it is handled under their terms. Nothing is sent before that press.
- The AssemblyAI helper. Only if you use AssemblyAI live: your AssemblyAI key is sent to a small helper running on this computer (127.0.0.1), which forwards it to AssemblyAI to get a short-lived token and does not store or log it. The audio then goes from the page straight to AssemblyAI. With the helper running, finished file transcripts are also deleted from your AssemblyAI account; without it they stay there until you delete them in their dashboard.
- Audio. By default the app does not save the audio of a live recording: it is turned into text on its way to the provider and then gone from this page. Two settings (Settings, Saving) change that. Record the audio of live sessions (off by default) saves the sound next to the transcript, in the same folder or browser storage, in pieces such as
Name-part1.webm; nothing is recorded while paused. Keep a copy of audio files I transcribe (on by default) saves the file you loaded next to its transcript. These audio files stay on your computer; they are deleted with the transcript from the Library. If you record people, make sure they know. - An app on this Mac (Source). With the ZumiScribe Local app, the page can listen to another app's sound, such as a Teams or Zoom call. The ZumiScribe Local app takes the sound and passes it to this page on the same computer (
127.0.0.1), and nowhere else; from there it goes to your transcriber like a microphone's. It works only while the page is listening, and macOS asks once to allow it (Screen Recording, which covers app sound). The people on the call should know they are being transcribed.
What leaves your computer
Only to the speech-to-text provider you pick in Settings, and only while you record or when you transcribe a file. Live recording streams the audio as it is spoken. Transcribing a whole file uploads that file. The text that comes back is shown on screen and saved as above.
| Provider | Where the audio goes |
|---|---|
| Soniox | stt-rt.soniox.com (live), api.soniox.com (files) |
| Modulate | platform.modulate.ai (live and files) |
| Deepgram | api.deepgram.com (live and files) |
| Speechmatics | eu.rt.speechmatics.com (live; a short-lived key is fetched from mp.speechmatics.com), asr.api.speechmatics.com (files) |
| Gladia | api.gladia.io (live and files) |
| Gemini | generativelanguage.googleapis.com (files only: the file is uploaded to Google's file service, used, then deleted) |
| OpenAI | api.openai.com (files only) |
| ElevenLabs | api.elevenlabs.io (files only) |
| Mistral Voxtral | api.mistral.ai (files only; the attendee names and auto-correction words are sent with it as hints) |
| AssemblyAI | api.assemblyai.com (files; the transcript stays in your AssemblyAI account unless the local helper deletes it) and streaming.assemblyai.com (live, with a token fetched through the helper on this computer) |
| Azure Speech | Your own Azure Speech resource, <region>.api.cognitive.microsoft.com or the endpoint you set (files only). The audio stays in your Azure tenant under your own Azure terms; the attendee names and auto-correction words are sent with it as a phrase list. |
| Voz + Who (in this browser) | Nowhere: the audio is transcribed inside this browser. If Voz + Who is installed in the ZumiScribe Local app, the audio also goes to that app on this computer (127.0.0.1) for the speaker labels. A usage count goes to Desert Ant Labs (below). |
| Voz + Who (on this Mac) | Nowhere: the file goes to the ZumiScribe Local app on this computer (127.0.0.1:8768). A usage count goes to Desert Ant Labs (below). |
| Granite (in this browser) | Nowhere: the audio is transcribed inside this browser, and Granite reports nothing to anyone. Only with speaker labels turned on (Settings, Transcriber): the speakers are labelled by Desert Ant Labs' Who, in the ZumiScribe Local app on this computer (127.0.0.1) when it runs or else in this browser, and Who sends its usage count (below). |
| Granite (on this Mac) | Nowhere: the file or the live audio goes to the ZumiScribe Local app on this computer (127.0.0.1:8768), and Granite reports nothing to anyone. Only with speaker labels turned on: Who labels the speakers in the same app and sends its usage count (below). |
| Local engine | Nowhere: it runs on your own computer, at the address you set (by default 127.0.0.1). Nothing leaves it. |
Parakeet on an Android phone. With the ZumiScribe Local Android app and the browser bridge switched on, the browser sends the microphone's sound (or a file you choose) to the app on the same phone, over the phone's own address (127.0.0.1), and the words and speaker labels come back the same way. Nothing goes to the internet except the one-time download of the models (from huggingface.co and github.com), and the app reports nothing to anyone. The bridge answers only pages of ZumiScribe.
Voz + Who and Desert Ant Labs. The two Voz + Who transcribers use Desert Ant Labs' Voz (speech) and Who (speakers) models, under their source-available licence. Your audio, transcript and file names stay on your computer: in this browser, or in the ZumiScribe Local app. In the browser, the first use downloads the models from huggingface.co and their runtime from cdn.jsdelivr.net and unpkg.com (about 450 MB, then kept); the ZumiScribe Local app downloads them once when you install Voz + Who in its setup window. Once a month, each model tells Desert Ant Labs (events.desertant.com) that it was used, so they can count the devices using it: the request carries a random ID for this browser (or for this Mac, in the app), the site's or app's name, the model and its version, and like any request it reaches them with your internet address. It cannot be switched off. In the browser the ID and the month's licence are kept in its storage, and clearing the site's data makes a new one; the ZumiScribe Local app keeps them in its own folder. Before the first use ZumiScribe shows this and asks you to confirm.
Granite. Granite (in this browser) uses IBM's Granite Speech model (granite-speech-5.0-470m-turboctc), which IBM publishes under the Apache 2.0 licence: free to use, nothing to accept, and it reports nothing. The first use downloads the model (about 320 MB) from IBM's pages on huggingface.co and its runtime from cdn.jsdelivr.net, then keeps them in this browser's storage for this site. Granite (on this Mac) runs the same model in the ZumiScribe Local app, which downloads it once from the same place when you add Granite in its setup window. After that your audio, transcript and file names never leave this browser. Speaker labels are on unless you turn them off: Desert Ant Labs' Who labels the speakers, with the usage count described above. Turned off, nothing is reported to anyone.
Only if you connect an assistant: the optional zumiscribe-mcp program lets an AI assistant you choose (Claude Desktop or Claude Code, for example) search and read your saved transcripts. It runs on your computer and is read-only, and it never reads your private notes, actions or pictures, but the passages and transcripts the assistant asks for are sent to that assistant and to whoever runs it. Do not connect it to a library you would not show them.
Only if you use it: in the setup dialog, Read a screenshot reads a pasted picture of a meeting to fill in its title and people. If your browser has no built-in text reader switched on, it asks first and then downloads a small text-reading tool (Tesseract.js, about 10 MB) from cdn.jsdelivr.net and its English language data from tessdata.projectnaptha.com. Only those downloads use the network: the screenshot itself is read on your computer and is not sent anywhere.
Only if you set it up: Settings, Sharing, "Send to a web address" posts the finished transcript (as Markdown, text, data without your private notes and pictures, or a Slack message) to the address you type, once when you press Send or, if you switch it on, each time you stop recording. That address is wherever you chose to send it, and ZumiScribe does not control what it does with the text. The address is kept in this browser only.
The page itself comes from the web server that hosts it, which sees an ordinary page request, and it loads two fonts from Google Fonts, which see the same. The app adds no analytics or trackers of its own.
What you are responsible for
Check each provider's terms. How long a provider keeps audio or text, and whether it uses either to improve its models, is set by that provider and by the plan you are on. It has not been checked here and it can change. Read the terms and privacy page on the provider's own site (Soniox, Modulate, Deepgram, Speechmatics, Gladia) before recording anything sensitive, and use the local engine when the audio must not leave your computer.
- Tell people. Laws and workplace rules about recording and transcribing conversations differ between countries, states and employers. Some need everyone's consent. A transcript of people's words is personal information. The Announce button speaks a notice, and a line can show on screen while you record (Settings, Notices), but a notice helps people give consent; it does not replace it where the law requires it. This page is not legal advice.
- Names and voices. People are named by hand. The app does not recognise anyone by their voice; the speech services separate speakers within one recording and that is all.
Keeping less, and deleting
- Delete a transcript from the Library (it is removed from the folder or the browser storage for good), or delete the file yourself in your folder.
- Clearing this site's data in your browser removes the keys, the settings and every transcript kept in the browser.
- Settings, Privacy has three tools for this: Erase everything on this computer (one confirmed button that removes every key, setting, saved list and recording kept in the browser; files in your chosen folder are not touched), Keeping recordings (delete ZumiScribe's own transcripts and their audio after 1, 7, 30 or 90 days or a year; it never touches other files and never runs while recording), and Anonymous mode (never asks who a voice is, never takes a name from speech, remembers no names or attendee lists).
- Pictures, files and links you attach to a bubble are kept only inside the transcript file (a picture is shrunk and stored in it; a dropped file is only a name and size, not a copy). They are not in the text exports and are not sent to phones that follow along, and they go to nobody unless you share the .json file yourself. A recording kept in a folder of its own keeps its full pictures in that folder, and a PDF you make from the export menu contains the pictures only if you tick them (your notes are off by default).