ZumiScribe Open ZumiScribe
‹ Back to the ZumiScribe home page Open the app

Questions and answers

Something not working? Check my setup finds the common causes, or see If something goes wrong. The common questions, with short answers. The ZumiScribe Local app page and the data handling page go deeper.

Basics Privacy and your data Speech services, the local engine, Voz + Who and Granite Using it Recording, audio and long recordings Voices, notes and captions If something goes wrong The ZumiScribe Local app and servers

Basics

What is ZumiScribe?

A web page that listens, writes down what is said as it happens, and marks who said it. It is meant for meetings, calls, panels and talks, and for audio and video files you already have. It can also just capture the sound to transcribe later, queue many recordings in a batch, and run one recording through several transcribers so you can compare them. Your recordings stay on your computer. It has no server or account of its own.

Do I need to install anything?

No. In Chrome or Edge, add a key from one speech service in Settings, press Record, and you are working. Or choose Granite (in this browser) or Voz + Who (in this browser), which need no key at all (see below). The optional ZumiScribe Local app adds private on-Mac transcription, captions followed on phones, AssemblyAI live, your company glossary and an assistant connector.

Which browsers work?

Chrome and Edge on a computer work fully, including saving into a folder you choose. Other browsers can transcribe but keep recordings inside the browser instead of in a folder, because they cannot write to a folder on your computer. Choosing a folder of recordings in Batch also needs Chrome or Edge.

Is it free?

ZumiScribe itself has no account or subscription. Most cloud speech services have a free trial or a number of free tokens or hours each month, then bill you directly, usually per hour of audio. Prices change, so check each service's own page (the welcome lists where to get a key for each). The local engine costs nothing to run but takes about 6 GB of disk. Voz + Who and Granite are free too, and run in your browser with no key.

Can I try it without setting anything up?

Yes. Try a sample recording first in the welcome lets you pick one of four public-domain readings (Anne of Green Gables, The She-Wolf, A Scandal in Bohemia and The Valley of the Shadow of Death, from LibriVox). It plays as a live session, at the speed of the recording, so you see the speaker colours, the captions and the controls. Nothing is downloaded to your computer. It does go through a speech engine like any recording, so pick one first: a cloud service usually has a free trial, and Voz + Who or Granite in this browser need no key.

What does the Chrome built-in AI option need?

It is a small model that desktop Chrome 138 or later can run for summaries. There is nothing to install in ZumiScribe, but Chrome has to have the model: it downloads it once (a few GB), needs roughly 22 GB of free disk and a recent graphics card or 16 GB of memory, and may remove it again when disk space is low. Some computers cannot run it at all, and Chrome's own page chrome://on-device-internals says why. It is best for short and medium meetings in English, Spanish or Japanese. A long recording is read in pieces, so the result is less detailed than from Ollama or a cloud model.

Privacy and your data

Where are my recordings stored?

In a folder you choose (by default each recording gets its own subfolder with the transcript, audio and pictures, as ordinary files; a setting under Saving keeps them all in one folder instead), or, if you skip that, inside this browser. The browser's own storage can be cleared, so a folder is safer. You can move recordings into a folder later. Recordings are written as you go, so a crash or a closed tab loses very little.

Is my audio sent anywhere?

It depends on the speech engine you pick. With a cloud service the audio goes to that company while you record. With the local engine (the ZumiScribe Local app plus its speech engine) nothing leaves your Mac. With Voz + Who (and Local (Parakeet + Who), which uses Who for the speakers) your audio stays on your computer, but the models send Desert Ant Labs a usage count once a month (no audio or text). What happens to your data lists it service by service.

Are my notes and attachments shared?

Your Notes panel, the notes, actions and bookmarks on bubbles, and pictures and files you attach are private: they are never in exports (text, Markdown, subtitles, copy), never shown to phones that follow along, never sent to a web address, and never read by the assistant connector. They are saved with the recording on your computer. Two of them are used when you ask for a summary: a note or action on a bubble is given to the summary model with the words, so that an action you noted appears in the action items (a cloud summary model, if you chose one, therefore receives them too, and asks before sending anything). The Notes panel text, bookmarks and attachments are not given to the summary.

Is it safe to put an API key in the browser?

The key stays in this browser. In Settings you choose how: until the tab closes, on this device, or on this device locked with a passphrase. Plain storage can be read by anyone using the same browser profile, so use the passphrase option on a shared computer.

Speech services, the local engine, Voz + Who and Granite

Which speech service should I choose?

Step 2 of the welcome has a table of every service: whether it works live or on files only, whether it labels speakers, and where to get a key. Soniox, Modulate, Deepgram and Speechmatics work live with speaker labels. Gladia works live but labels speakers on files only. Gemini, OpenAI, ElevenLabs, Mistral and Azure work on files you load, and AssemblyAI works live with the ZumiScribe Local app. Voz + Who works live and on files with no key, in this browser or in the ZumiScribe Local app, and so does Granite, in this browser. On a Mac, Local (Parakeet + Who) uses the Parakeet engine for the words and Who for the speakers, which is much faster than Local (Parakeet) alone. Most of the live services can also transcribe a whole file, which is much faster than playing it live. In Settings, Transcriber, each one shows what to know about it: where the audio goes, what it needs and its limits. How they compared on one recording may help.

What is the local engine, and what does it need?

A speech engine (Parakeet) that the ZumiScribe Local app installs and runs on your Mac. It needs an Apple silicon Mac (M1 or newer), about 8 GB of memory and about 6 GB of free disk. The first install takes a quarter of an hour or more. It is offered two ways in Settings, Transcriber: Local (Parakeet), which also labels the speakers itself (slow), and Local (Parakeet + Who), which only writes the words and has Who label the speakers (fast).

Local (Parakeet) or Local (Parakeet + Who)?

Both use the same Parakeet engine on your Mac. Local (Parakeet) also works out the speakers inside the engine (up to 4 voices), which is the slow part. Local (Parakeet + Who) switches that step off, so the engine only writes the words, which is fast, and then Desert Ant Labs' Who labels the speakers: automatically for recordings you load (including in Batch), and with the Who button after a live recording (keep the audio for that). Who runs in the ZumiScribe Local app when it is installed (setup step 5), which is quicker, or in this browser. Who's usage count to Desert Ant Labs applies, as for Voz + Who. If you only want a quick transcript with no speakers, either works; if you want the fastest labelled transcript on your Mac, pick the second.

Where does the engine keep its files, and how do I remove it?

In ~/Library/Application Support/ZumiScribe/engine. In the app's setup window, Show the folder opens it and Remove it deletes it after asking. Your recordings and glossary are not touched.

What is Voz + Who?

Two models by Desert Ant Labs: Voz writes down the words (it is NVIDIA's Parakeet model in their runtime) and Who works out who is speaking. ZumiScribe runs them on your computer, with no account and no key, in one of two places:

It understands 25 European languages and finds the language by itself; for other languages it writes nonsense, so the Spoken Language setting must be one of them or Auto-detect. Live, the newest words can change for a second or two before they settle, and labels for a voice that only speaks briefly can be split between numbers. It never runs on a company server: its licence is for your own device.

What does Desert Ant Labs see?

Not your audio, transcript or file names. Once a month each model tells them it was used, with a random ID for this browser (or this Mac), the site's or app's name and the model version, so they can count devices: it is free below 100,000 devices a month per model, and paid above. This cannot be turned off; ZumiScribe asks you to confirm it before the first use. The details.

What is Granite?

IBM's Granite Speech model (granite-speech-5.0-470m-turboctc), free under the Apache 2.0 licence. Granite (in this browser) runs it inside this browser, live or on files, with no account, no key and nothing to install. It needs Chrome or Edge with WebGPU on a computer, and downloads about 320 MB the first time (kept by the browser after that).

Granite (on this Mac) runs the same model in the ZumiScribe Local app (setup step 5, about 320 MB more): it works in any browser, including Safari and Firefox, files and live.

It does not tell speakers apart by itself. Turn on Label speakers under it in Settings, Transcriber, and Desert Ant Labs' Who labels them: on files in this browser or, faster, in the ZumiScribe Local app; live only with the ZumiScribe Local app. With that on, Who's monthly usage count applies (below).

Granite or Voz + Who?

Both run on your computer with no key. Granite needs nothing from anyone and was as accurate as Voz on our test recordings, and quicker; it covers 7 languages to Voz's 25, and needs Chrome or Edge. Voz + Who labels speakers itself and works in Safari in the browser; both run in the ZumiScribe Local app too. With speaker labels on, Granite uses Who for them, so the difference there is small.

Does it work offline?

The page loads without internet after the first visit. Transcribing needs a connection for a cloud service; the local engine needs none once it is installed, and Voz + Who and Granite work offline once their models are downloaded (Voz + Who reports its usage when it is next online). Recording from the microphone with Voz + Who has been tried with the network cable out. The first download needs a connection; cloud services need one throughout; and Ollama or Chrome's built-in AI can write summaries, and a downloaded language can translate, without one.

Using it

How does it know who is speaking?

The speech engine separates the voices and gives each a colour. You name them: click a bubble, then a name from the Voices panel, or add the attendee list before you start and let introductions do it. After the meeting, Review plays a few bubbles of each voice so you can name them or move bubbles, Check steps through the places a label is probably wrong, and Who labels everything again from the sound.

Can I transcribe a recording I already have?

Yes. The Transcribe a recording tab takes an audio or video file (mp3, m4a, wav, mp4, mov, webm and similar; only the sound is used) and returns a transcript with speakers marked. A service that handles whole files does it in a fraction of the recording's length, or you can play it as a live session with a live service. You can also paste a link to an audio file: the web page that holds it must allow other sites to read it (archive.org does; many others do not). If a link is refused, download the file and use "Audio or video file" instead. Only use recordings you may use: public-domain readings such as LibriVox are fine, while podcasts and broadcasts are usually copyright.

Can I transcribe a whole folder, or many recordings at once?

Yes, in the Batch tab. Add files by dropping them in, choosing them, or Add a folder... (Chrome and Edge on a computer; it takes the audio and video files directly in that folder, not those in folders inside it, and leaves unticked any that already have a transcript beside them). Tick a transcriber, press Run, and the files are transcribed in a queue: up to two at a time on cloud services, one at a time on engines that run on your computer. You can pause, cancel one job or all of them, and retry any that failed; one failure does not stop the rest. Each result is saved as its own transcript in your recordings folder and opens from the queue. A box above Run, Keep a copy of each recording's sound beside its transcript (on unless you untick it), decides whether the sound is copied there too: without it the bubbles have no play buttons and Who cannot be used afterwards. A cloud service uses the credit or tokens of your key, and the plan line shows an estimate before you start. The queue runs while the page is open, so keep the tab open.

Can I compare transcribers on the same recording?

Yes. In Batch, add the recording and tick two or more transcribers (the list shows every service that works on whole files). Each combination is one job, and each result is its own transcript, named like Talk - Soniox, all sharing one copy of the sound. When a file has finished with two or more transcribers, Compare opens a table (words, speakers found, time taken, estimated cost) and the texts side by side, with the words that differ from a chosen reference marked and the share of words that differ shown for each. It shows how much they disagree, not which is right, unless the reference was corrected by hand. No service reports what it billed, so the cost is an estimate from the price per hour you enter in Settings, Transcriber.

Can it transcribe phone and video calls?

For calls on your computer, yes: in the start steps choose Online meeting (or Phone call for two people on speaker) and an audio source that captures the sound of a browser tab or the screen (Chrome and Edge). With the ZumiScribe Local app, An app on this Mac listens to the Teams or Zoom app itself. It cannot listen to calls on a phone itself.

Why does a film or TV app go black while ZumiScribe listens to it?

Apps that play protected video, such as the Apple TV app (and Netflix or Disney+ in Safari), black out their picture whenever macOS is capturing the screen, so that it cannot be copied. On a Mac an app's sound can only be taken through a screen capture, so listening to such an app turns its picture black. The sound still comes through and is transcribed, but snapshots of it are black too. This is the app's copy protection and cannot be switched off. Meetings, webinars and slides are not affected.

Can an Android phone transcribe privately?

In an early test version, yes: the ZumiScribe Local Android app runs Parakeet and speaker labels on the phone, and ZumiScribe in the phone's browser uses it (Settings, Transcriber, Parakeet (on this phone)). The sound goes from the browser to the app on the same phone and nowhere else. It needs a recent phone, about 680 MB of models, and warms the phone and uses its battery on long meetings.

Can the audience follow the captions on their phones?

Yes, with the ZumiScribe Local app. Turn on sharing on your network, then open the address or scan the code on a phone on the same Wi-Fi. Nothing needs installing on the phones. The same captions can fill a big screen from this computer: View, Captions.

What can I do with a recording afterwards?

Edit it, search every meeting, ask questions across all of them, review and fix the speakers, write a summary with decisions and action items, and export to Markdown, PDF or Word with the pictures. You can also email a follow-up summary. Every recording is listed in the Library, where you can rename, download or delete it.

Do summaries need anything extra?

Yes, a text model: Ollama on your Mac (private, free), the small model built into desktop Chrome (private, but Chrome has to have it: see above), or a key for OpenAI or Gemini (both have free allowances). Skip it and everything else still works. Cloud summaries ask before the transcript text is sent.

Can it translate?

Yes, in Chrome 138 or later on a computer. The Translate button shows each bubble's translation under the words, and the large captions have their own From and Translate to menus. A language is downloaded once and the translating is done on your computer. Change From at any time if the talk switches language.

Are translations saved, and are the original words kept?

The original words are never replaced. Finished sentences are saved in the recording's file beside them and used again later. Exports ask which language to use when a recording has translations.

The slides in my webinar keep being captured when nobody changed slide. Why?

The people, cameras and chat change all the time and can look like a new slide. Choose a capture area (step 4 of the start steps, or the Area button during a recording) so only the part with the slides is watched. The picture that is kept is still the whole screen.

Recording, audio and long recordings

Can I just record the sound and transcribe it later?

Yes. Press Audio only beside Start recording (or in the toolbar). Nothing is transcribed while you record, so it needs no signal, no key and no tokens: useful on a train, in a room with no wifi, or when your credit has run out. A recorder screen shows the sound against the timeline (the bars light up where speech is heard), a timer, and the files as they are saved. When you press Stop, Transcribe is there at once, and the recording is also listed in the Library under Audio waiting to be transcribed, with its own Transcribe button, so you can do it later with any engine.

Can I pause a recording?

Yes, for both kinds. Nothing is recorded while paused, which suits off-the-record moments or a stretch you do not want. On the audio recording a pause shows as a grey gap on the timeline, and the next piece starts as a new file. A live transcription that stays paused for 15 minutes is closed; an audio-only recording is not.

How do I keep the sound as well as the transcript?

Choose Record and keep the audio in the Record menu (or Audio and transcript in step 4 of the start steps), and tick Remember this choice to make it the default. The sound is saved beside the transcript. You then hear any bubble by pressing play on it, and Who can label the speakers. Without the sound there is nothing to play back, and Who cannot be used.

How often is the audio saved, and what are the "parts"?

A recording is cut into files of up to a chosen length (one hour unless you change Longest single audio file under Settings, Saving), named Title-part1.webm, Title-part2.webm and so on, and the file being recorded is saved every 30 seconds, so a crash or a closed tab loses at most the last moments. A pause also starts a new part. If there is no folder or browser storage, the sound is held in memory until you stop: ZumiScribe warns you, and a folder is safer.

What happens if I refresh the page or close the tab while recording?

The browser asks first, and recording stops if you go ahead, because the microphone belongs to the page. What was already saved is kept: the audio files, and the transcript up to the last moments. Open ZumiScribe again and the bar at the top offers to Restore it.

Can I record a very long meeting?

Yes. Save into a folder so the audio is written as you go, keep the screen awake (ZumiScribe asks the browser to), and use a live service or the local engine. After a long recording, Who and the summary can take a few minutes. Summaries read a long transcript in pieces, so the result may be less detailed with the Chrome built-in model than with Ollama or a cloud model.

Voices, notes and captions

How do I fix the speakers afterwards?

Open the Voices panel (it sits at the right). Its top has three tools: Review shows a card per voice with a few sample bubbles you can play, then lets you rename the voice, give it to another speaker, or give single bubbles to someone else, with Undo; Check steps through the places where a label is probably wrong; Who listens to the sound and labels everything again. You can also click a bubble and then a name.

If I press Who, are the speakers I assigned replaced?

Yes. Who listens to the whole recording and gives each sentence the voice that speaks it, so every speaker you assigned, including names you gave to voices and bubbles you moved by hand, is replaced. ZumiScribe warns you before it starts. Notes and bookmarks stay with the first part of their bubble. If you want to keep your own labels, use Review or Check instead, or save a copy first. Who needs the sound to have been kept, and needs recording to be stopped.

What is the Notes panel, and how is it different from notes on a bubble?

Annotate, Notes panel opens a side panel for your own notes on the whole recording, with a paragraph style (three heading levels), bold, italic, underline, and bulleted and numbered lists. It is not tied to a bubble or a time. It can stay open next to the Voices panel. Bubble notes (Annotate, Bubble notes, or Alt+N) are small notes and bookmarks attached to a particular bubble. Both are saved with the recording and are private.

Can I show large captions on a second screen?

Yes: View, Captions shows the last sentences in big type. You can set the text size, show earlier sentences and names, and turn on Colour so each speaker's words are in their own colour and a change of voice is clear. The same captions can go to phones on your Wi-Fi with the ZumiScribe Local app. The text size in the transcript has its own slider in the View menu.

If something goes wrong

Start here: Check my setup tests the microphone, the ZumiScribe Local app, the speech engines (Parakeet, Voz + Who, Granite, Qwen3-ASR, WhisperLiveKit), Ollama and the relay in one go, and can copy a report for a support message. The same check is in ZumiScribe under Settings, Services.

There is no signal, or my speech service has run out of credit

If you know before you start, use Audio only and transcribe later. If it happens mid-meeting, the recording that was kept as audio can still be transcribed afterwards from the Library (Transcribe a recording also takes any audio file). Top up the service, or switch to Voz + Who or Granite in this browser, which need no key and no connection after their first download.

Transcribe says it needs a transcriber for whole files

The Batch tab (and the Library's Transcribe all) works on whole files, which not every service offers. Open Settings, Transcriber and pick one that works on files, with its key set. Nearly every service in the table does (Soniox, Deepgram, Speechmatics, Gemini, OpenAI, Mistral, Voz + Who and Granite among them), so this only happens when none of them has a key yet. A single recording can still be played live into a live-only service.

Chrome shows an error about a microphone or sharing

Check that Chrome may use the microphone for this site (the icon at the left of the address bar), that no other app is holding it, and, for a shared tab or screen, that you ticked Also share tab audio in Chrome's picker. Check my setup tests the common causes.

A file says "Waiting for the engine" for a long time

The engine on your Mac (Local (Parakeet), Local (Parakeet + Who) and the ZumiScribe Local app's Voz + Who) works on one file at a time, so a file sent while another is being worked on waits its turn. Files with speaker labels switched on in the engine (Local (Parakeet)) can take far longer than the recording, which holds up everything behind them: a 30-minute file can take over twenty minutes. Local (Parakeet + Who) does not have that step, so a recording that takes minutes there can take seconds. If you reload or close the page while files are being worked on, they are left in the engine's queue; the Batch tab lists any it finds under Jobs left on the engine, with Cancel, and leaving a page now also tells the engine to drop its jobs (once the ZumiScribe Local app has been updated).

The play button is missing from the bubbles

The bubbles have a play button only when the recording's sound was kept and can be found beside the transcript. A recording made with Transcript only, or a Batch run with Keep a copy of each recording's sound unticked, has none, and it cannot be added afterwards: record or run it again with the sound kept. If the sound was kept and the button is still missing, reload the page; if that brings it back, please tell us how the recording was made.

The Run button in Batch is greyed out

Run needs at least one ticked file and at least one ticked transcriber that is ready (a service with no key yet cannot be ticked). Once you press Run the jobs are in the queue, and Run stays off until you change the files or the transcribers, so the same jobs are not queued twice.

My transcript from a link or a file is empty

The speech service found no words, usually because the language is not the one set under Spoken language, or the file has no speech. Set the language (or Auto-detect) and try again. For a link, make sure it points at the audio file itself, ending in .mp3, .m4a, .wav or .mp4, and not at a web page about it.

The ZumiScribe Local app and servers

macOS says it cannot verify the app. Is it safe?

The app is not signed with an Apple Developer ID yet, so macOS asks you to confirm it once. The steps are on the install page, or install with the Terminal command there to avoid the warning.

Chrome asks to allow access to devices on my network. Why?

Chrome calls any page that talks to a program on your own computer "access to other apps and services on this device". ZumiScribe asks only if you use the ZumiScribe Local app or the local engine, and only after the welcome has explained it: nothing on your computer is contacted before that. If you only use a cloud service or a speech engine in the browser, you can decline and nothing changes. If you use the ZumiScribe Local app, allow it, or the page cannot reach the app. Check my setup confirms that it works.

Can a whole company use it with nothing installed?

Yes: the same services can run on a company server. The server has no sign-in yet, so use it only on a trusted network. See the administrator guide and the three ways to run ZumiScribe.