Choosing a model: Which device and which model to transcribe with?

The three options

This device: Transcription happens on your own device with the model you choose, and the audio goes nowhere. The model downloads once in the background and goes into memory when you press Prepare the model. You change the choice in Settings on the transcript page.

Transcriber device: You make a strong computer of yours, or a teammate’s, a transcriber. Sign in there with the same account and press Make this device a transcriber under My account → Transcriber device. Your phone or weak computer sends it the audio peer to peer, and the audio does not pass through the server. A teammate’s device needs the transcript to be shared with the team.

Cloud transcription: The audio goes to the service you choose and is transcribed there by a large model. No model downloads to your device and no graphics card is needed. Accuracy is at about the level of M3 or a little below. Live, the audio goes in pieces of at most 30 seconds, so the text comes later. On finished files it is the fastest route, for files of up to 2 hours. Getting a free key is covered in free Groq API key.

What separates M1, M2, M3 and M4

* The rough share of words written correctly on Turkish FLEURS clips. Noise and overlapping speech lower it in real use. The method, and English results, are in our Whisper test.

On a phone, only M1 is available below 6 GB of memory, and M1 and M2 above it. Models that do not run on this device are marked in the list. The Status page in the top menu shows whether your device meets these needs, and browsers and devices has the details.

What LiveType picks for you

If you touch nothing, a computer with a graphics card gets M3, one without it but with 8 or more processor threads gets M2, and other devices get M1. Phones default to M1. You can change your preference with the Speed ↔ Accuracy setting on the My account page.

When the text falls behind, or the device could do more

During transcription LiveType measures the real speed. If the text falls more than 10 seconds behind the live audio, it suggests a smaller model. If your device can keep up with more, it suggests a larger one. The change applies from the next start, and the decision is yours.

When a file runs in the background, delay does not matter: on a slow device a large model takes longer but gives accurate text. Being able to edit while it transcribes also softens a small model’s mistakes: you fix each one as you see it instead of leaving them all for the end.

If the graphics card is not used at all, try the WebGPU steps in transcription troubleshooting. If the device simply cannot keep up, move to cloud transcription or a remote device.

Things to keep in mind

• The first download can be hundreds of megabytes or more, so prefer Wi-Fi.
• The model stays in your browser’s storage. Clear your browser data and it downloads again.
• A private window cannot keep the model.
• If keeping the audio private comes first, read private transcription before using the cloud option.

With 30 minutes on the free plan, you can compare models on your own recording. Plans are on the pricing page.