Why the same model runs at wildly different speeds
Speech recognition models do enormous numbers of simple calculations at once, which is precisely what a graphics card is built for and precisely what a general purpose processor is not. The same model and the same audio can differ by more than an order of magnitude between the two, which is why advice from someone with a desktop graphics card is useless to someone on a thin laptop.
Model size compounds it. The larger Whisper models are meaningfully more accurate on difficult audio and dramatically heavier to run, so the accuracy you want and the speed you want pull in opposite directions on hardware that has no graphics card to fall back on.
What actually helps, in order
Dropping to a smaller model is the change that makes the biggest difference, and on clear recordings the accuracy cost is smaller than people expect. Using a version of Whisper optimised for processors rather than the reference implementation is the next largest gain, often several times faster on identical hardware. Quantised models cut memory use and speed things up again.
There is a point where fighting this stops being worth it. If you transcribe occasionally and your laptop takes longer than the recording itself, the calculation is straightforward: uploading the file to hardware built for the job takes minutes rather than an afternoon, and your first file here costs nothing to compare against what you were getting locally.
Frequently asked questions
Would a smaller model fix this?
Usually, and it is the first thing to try. On clean, clearly spoken audio the smaller models lose less accuracy than their size difference suggests, while running many times faster on a processor.
Is running Whisper myself still worth doing?
Often yes, particularly if you transcribe constantly or the audio must never leave your machine. Our guide on running it yourself covers what the setup genuinely takes and where people get stuck.
What hardware do you run it on?
A machine we own with a graphics card, which is the whole reason it finishes quickly. That is also why the audio never goes to a third-party service: the processing happens on our own hardware.