Transcribe a meeting on your device, no upload

Why you’d do this

A recording of a meeting usually contains the things you least want on someone else’s server: names, figures, half-formed decisions, a client’s complaints. Most transcription services need the audio uploaded before they hand back a word. Dzzz’s Transcribe tool runs the speech model inside your browser instead, so the file is read and processed on your own device. You get a timestamped transcript you can play back, correct and export, and the audio never travels anywhere.

The trade-off is honest and worth knowing up front: the model has to be downloaded once, and the work is done by your phone or laptop, so a long recording takes real time. For a one-hour meeting on a modern computer that is a fair price. On an older phone it is slower, and the page tells you roughly how long to expect before you start.

How, with the named tool

1. Add the recording. Open Transcribe. On the Upload file tab, drop an audio or video file or tap to choose one. Alternatively, switch to the Record tab, record from your microphone, and tap Use this recording.
2. Choose a model. Under Model you will see three options: Base (balanced), Tiny (faster, less accurate) and Small (best Arabic).
   • Tiny is a 41 MB download.
   • Base is an 80 MB download and is the balanced default.
   • Small is a 249 MB download. It is noticeably better with Arabic, but much slower on phones, so the tool recommends a computer for it.
3. Set the language. Pick the language actually spoken in the recording from the Language list. This matters: choosing the wrong one is the commonest cause of poor output.
4. Download the model. Tap Download engine; its label carries the size of the model you picked. The download happens once and the model stays cached, so the next visit goes straight to the run button.
5. Transcribe. Tap Transcribe and keep the tab open. The page shows an estimate of the processing time and then the elapsed time. When it finishes, you get timestamped lines.
6. Edit and export. Tap a timestamp to play that line, correct the text in place, then export TXT, SRT or VTT, or use Copy transcript.

If your device supports WebGPU, the model uses it. If not, the tool falls back to a slower compatibility engine and says so while it works (“slower here because this device runs the compatibility engine”). The result is the same kind of transcript, just later.

Turning the transcript into a summary

Meeting Summary is the companion tool. It runs the same on-device transcription (the Base model, an 80 MB one-time download), then shows the transcript in an editable box. Read it and delete anything sensitive before you do anything else.

From there you have two choices:

• Basic on-device summary instead keeps everything local. It is an extractive summary: it picks sentences that are already in the transcript rather than writing new ones.
• Send for AI summary, where it is offered, sends the transcript to our server for a better summary. This is a separate, explicit tap, and the page labels it clearly: “This step leaves your device.” Our server does not keep or log the transcript, but it passes it on to an outside AI provider to write the summary, and we do not control how that provider handles it, so treat it as leaving your control. Nothing is sent automatically, and transcription itself never leaves your device.

What to watch for

Arabic accuracy depends heavily on the audio. The tool says this plainly: clear, close-to-standard speech transcribes reasonably well, while dialect, background noise or phone-call quality come out much rougher. Expect to correct the result, and plan time for it. Small is the best tier for Arabic, and if the output comes back in Latin letters instead of Arabic script, the audio was probably hard to make out, so try Small or check the language setting.

Meeting recordings are often the worst case. Several people talking over each other, a laptop microphone across the room and an air conditioner in the background all hurt accuracy in any language. A transcript of a meeting is a first draft, not a record.

Keep the tab open and the screen on. All the work happens in the page. On a phone, switching apps or letting the screen lock can interrupt a long run.

Sending the summary is the only step that leaves your device. Before tapping Send for AI summary, re-read the transcript box and remove names, numbers or anything you would not want on a server, even one that promises not to keep it. If in doubt, use the basic on-device summary.

FAQ

Which model should I pick?

Start with Base (80 MB) for a balance of speed and accuracy. Pick Tiny (41 MB) if your device is slow or storage is tight and you can accept more mistakes. Pick Small (249 MB) for Arabic, ideally on a computer. Each model downloads once and is cached independently.

Does the audio ever get uploaded?

Not by Transcribe, and not by the transcription step of Meeting Summary. The only thing that can leave your device is the transcript text, and only if you tap Send for AI summary yourself.

Why is it slow on my phone?

Everything runs on your own hardware. Without WebGPU the tool uses a slower compatibility engine, and the Small model is much slower on phones than on a computer. A shorter clip, the Tiny model, or a laptop will help.