Can Whisper run without a GPU?

Yes. Whisper runs on a processor with no graphics card, and you do not need a GPU to get a usable transcript. The trade-off is time. A GPU is usually the fastest route, but a CPU is usable: whisper.cpp finished a 13-minute recording in about two minutes with the small model ¹. Choose the implementation first, then the model size, then decide whether the wait is acceptable.

The processor route works, and the implementation changes the speed

whisper.cpp lists CPU-only inference among its features ². So a computer with no usable graphics card can still run it on its processor.

The CPU benchmark we found comes from the faster-whisper README. It transcribed a 13-minute recording ³ on a CPU using the small model. whisper.cpp took 125 seconds, or about two minutes ¹. The openai/whisper package took 418 seconds, nearly seven minutes, in fp32 precision ⁴. faster-whisper in int8 mode took 102 seconds ⁵, the quickest of the three.

These numbers come from one project's README, on one recording and one set of hardware. Treat them as the ranking, not as a promise for your computer. The ranking matters more than the exact seconds: openai/whisper is the slowest of the three on a CPU, and the two faster routes are both good choices. Our guide to running Whisper locally gives the install steps for each.

A GPU matters most for the large models and for speed

On a GPU, video memory decides which models you can run at all. Turbo needs about 6 GB of video memory ⁶. The faster-whisper README also measured GPU memory use. On an NVIDIA RTX 3070 Ti, faster-whisper in int8 precision used 2926 MB with the large-v2 model ⁷. That run was on one card, so read the figure as the scale of memory involved, not as a minimum.

On a GPU, faster-whisper needs an NVIDIA card with enough memory

For faster-whisper, the GPU route is NVIDIA-only in practice. Running it on a GPU requires NVIDIA libraries to be installed first ⁸. An NVIDIA card with at least 6 GB of video memory covers turbo ⁶. Check the video memory in your computer's graphics settings before you buy or download anything.

If your card is AMD, or your computer uses integrated graphics, faster-whisper's GPU route does not apply. Time the CPU route on the same test file before you decide anything else.

It depends on the model, the recording length and your patience

Three things decide whether the processor is good enough.

• Model size. The small model is practical on a CPU. The benchmark above does not time the large model on a CPU, so measure it yourself before you rely on it.
• Recording length. A short clip finishes quickly on any route. A long file is where a GPU saves real time.
• Batch work. If you transcribe many files each week, a GPU pays back its cost in time. If you transcribe one file now and then, the CPU is enough.

The processor mainly changes speed and memory. Whisper's accuracy still depends on the model, the language and the audio, so measure the word error rate on your own recordings before you judge a model.

What to do: test one short file on the CPU first

1. Check your memory. Turbo needs about 6 GB of video memory on a GPU ⁶, so a computer with less than that has to use the processor or a smaller model.
2. Pick the route. On a Mac, or on a Windows or Linux computer with no NVIDIA card, start with whisper.cpp. For Python work, use faster-whisper in int8 mode.
3. Run a five-minute test file with the small model, and time it.
4. Decide. If the wait is acceptable, keep the CPU route. If not, use a GPU, or move to a smaller model.

Before you choose, read our answer on which Whisper model to use, because the model size sets the memory you need. For the difference between the two faster CPU routes, see whisper.cpp versus faster-whisper. If you have no computer capable of running Whisper, running Whisper in Google Colab gives you a notebook with a GPU, within its free-tier limits.

For the full subject, see the speech recognition overview and the guide to Whisper AI.