Transcription troubleshooting: No text, slow, WebGPU and common messages

Slow transcription: Is WebGPU on?

In the browser, transcription is fastest on a graphics card, and WebGPU is what makes that possible. Without it the processor does the work: Text falls behind and accuracy drops. The transcript page shows its own notice in that case and tells you what to do for your browser.

To check yourself, type chrome://gpu in the address bar (edge://gpu in Edge). If the WebGPU line reads “Hardware accelerated”, the card is in use. If not, go through these in order:

1. Update your browser. New versions turn WebGPU on by themselves for many cards.
2. Turn on hardware acceleration. At chrome://settings/system in Chrome or edge://settings/system in Edge, turn on graphics or hardware acceleration and restart the browser.
3. Choose the right card. See the Windows or Linux steps below.

In Firefox and Safari it depends on the version and system: See browsers and devices.

Choosing a graphics card on Windows

Laptops may have two cards, and the browser may use the weaker one. On Windows 11 go to Settings → System → Display → Graphics. If your browser is not listed, add it as a desktop app. Click its row, set GPU preference to High performance, then close the browser fully and open it again. On Windows 10 the same setting is under Display → Graphics settings.

WebGPU on Linux

On Linux the card often needs extra settings. NVIDIA cards need Vulkan for WebGPU, but Vulkan under Wayland can show videos white. In that case LiveType shows a command on the page for your distribution and browser, which opens the app in its own profile with the right settings. Use the command in the notice rather than trying flags by hand.

Text falling behind: A smaller model

If transcription falls clearly behind the speech, the page suggests a faster model, and Switch to this model changes it. If you have room to spare, it suggests the opposite. If it still cannot keep up, you have three options: Use a smaller model, close other tabs, or let the cloud or a stronger device do the work. Choosing a model explains the models, and AI transcription accuracy explains why quality drops.

Falling behind is no problem with a file: The background mode works at your device’s pace and simply takes longer.

Common messages and their fixes

The model will not download, or downloads again

Browsers may delete model files when space runs low. If you see a storage notice, bookmark the site or install it as an app, and storage becomes lasting. Private windows cannot do that, so open a normal one.

• With data saving on, the model waits for your click rather than downloading by itself. Use Wi-Fi if you can.
• On a company network, Hugging Face may be blocked, and then the model never arrives. Try another network or move to the cloud.
• In an in-app browser, large downloads and the microphone may be blocked. Choose Open in browser from its menu.

Problems on a phone

Phones cannot share tab audio, so stream platforms do not work there and need a computer. Microphone, sound files and streams that allow it do work. Only small models run on a phone. For long jobs, plug the device in and keep the screen on, since the browser may stop the work when the screen turns off.

Still stuck?

Send the exact message, your browser, your device and the source through the contact page. You need not send the audio; most problems are clear from those details.