Private transcriptionHow to transcribe a recording without uploading it

Why most transcription uploads your audio

Speech recognition used to need a large model on a powerful server, so almost every transcription service works the same way: you upload the recording, their servers transcribe it, and you get the text back. That is true of meeting recorders, dictation apps and upload-and-wait sites alike. What happens to the audio afterwards (how long it is kept, who can hear it, whether it is used to improve their models) is then a matter of their policy.

That matters when the recording is somebody else’s words: an interview, a client call, a therapy or medical note, an HR meeting, a research participant. Often the promise you made to that person, or the rules you work under, is that the recording goes nowhere.

How in-browser transcription works

Small speech models are now good enough to run on an ordinary laptop, and browsers can run them.

1. The model comes to you. The first time, the page downloads a speech model (in our case Whisper base, about 80 MB) and your browser keeps it. Nothing is installed.
2. The audio is read in the tab. When you drop in a file, the browser decodes it and the model transcribes it on your device. A recording made in the page is transcribed as you speak, a sentence or so behind.
3. The text is yours to use. Timestamped lines, a search box over them, copy, and download as a .txt file.

How to check that a tool really keeps your audio on your device

“Private” on a website is a claim. You can test this one yourself in a couple of minutes:

• Watch the network. Open your browser’s developer tools (F12, or Cmd-Option-I on a Mac), choose the Network tab, then transcribe a file. The first time, you’ll see the model files load. You should not see any request large enough to carry your recording, or any that goes to a transcription service.
• Try it with the network off. Transcribe one short file with the page open. Then turn off your Wi-Fi and, without reloading the page, transcribe another. If it still works, the audio cannot have gone anywhere. (Reloading the page needs the connection again, because the page itself comes from the web.)

On Zero Witness the network tab will also show small page-view signals sent to our own site. They say that somebody is using the page. They carry nothing you recorded or typed.

How to transcribe a recording with Zero Witness

1. Open the transcriber.
2. Drop an audio file onto the page, or press the microphone to record. Anything your browser can play works: mp3, m4a, wav, webm, ogg, flac.
3. If you know the language, name it. It’s detected otherwise, but naming it avoids the one bad failure: a wrong guess produces fluent text in the wrong language.
4. The first time, the model loads (about 80 MB). After that it starts at once.
5. Copy the transcript, download it as .txt, or keep it in your browser.

What it does well, and what it doesn’t

• Best on clear speech, one speaker, in English. Whisper base is a small model. Accents, people talking over each other and background noise make it worse, and quality falls off outside English.
• No speaker labels. You get timestamped lines of what was said, not who said them.
• Unclear passages are marked, not guessed. Where the model loops or returns nothing, the line reads [unclear]. That is a rule about what the model did, not a confidence score, so it will not catch a fluent wrong word. Check anything you rely on against the recording.
• Speed depends on your device. On a recent laptop a minute of audio takes roughly ten to fifteen seconds. An older machine or a phone is slower.
• No limits on length or minutes. Files up to 200 MB, because the decoded audio is held in memory while it is read.

More on the transcriber itself: private audio transcription. Related guides: How to run an AI model in your own browser · How to tell whether an AI chat is really private · How to turn a photo into a PDF without uploading it · How to get a deposition summary from AI that you can check. All guides.