Can Whisper transcribe without a GPU? What a CPU can do, from a public benchmark, and where it falls short.
Short answer. Yes. Whisper runs on an ordinary CPU, more slowly than on a GPU and best with the smaller models. In the faster-whisper project’s benchmark, the small model transcribed 13 minutes of audio in between 51 seconds and 6 minutes 58 seconds on an Intel Core i7-12700K using 8 threads, depending on the program and settings, and used 1.0 to 4.2 GB of RAM. The fastest CPU setting was 8-bit with batching. The large model was not tested on a CPU in that benchmark, and smaller models are less accurate, so check Hindi and Hinglish recordings before relying on a CPU-only setup.
01
Measured on a CPU
These are the faster-whisper project’s own measurements on an Intel Core i7-12700K with 8 threads. The last column divides the 13-minute length of the audio by the time taken, my arithmetic. Even the slowest setting finished in under 7 minutes, which is 1.9 times faster than the recording lasts.
Program and precision (small model, beam size 5)
Time for 13 minutes of audio
RAM
Faster than real time
openai/whisper, 32-bit
6 min 58 s
2,335 MB
1.9 times
whisper.cpp, 32-bit
2 min 05 s
1,049 MB
6.2 times
whisper.cpp with OpenVINO, 32-bit
1 min 45 s
1,642 MB
7.4 times
faster-whisper, 32-bit
2 min 37 s
2,257 MB
5.0 times
faster-whisper, batch size 8, 32-bit
1 min 06 s
4,230 MB
11.8 times
faster-whisper, 8-bit
1 min 42 s
1,477 MB
7.6 times
faster-whisper, batch size 8, 8-bit
51 s
3,608 MB
15.3 times
02
What this tells you
Software and settings mattered a great deal. The same model and the same CPU took 6 min 58 s with the original code and 51 s with faster-whisper at 8-bit with batching, about 8 times quicker.
Memory is modest: 1.0 to 4.2 GB of RAM for the small model. Batching raises speed and memory together.
The small model is not the large model. OpenAI’s table puts large at about 4 times slower than small and about 10 GB against about 2 GB of VRAM, and the project says accuracy varies widely by language, so a CPU result for small does not transfer to large-v3.
03
Where a CPU falls short
Accuracy in Hindi, Hinglish or noisy rooms, which usually needs a larger model than small.
Volume. Many hours of recordings a day, or several meetings finishing at once, queue up on a CPU.
Sharing. The same processor often runs the language model and everything else, so transcription competes with chat.
Live captions. The figures here are for finished recordings, not streaming.
04
A practical rule
A few recordings a day, handled in the background, can run on a CPU with a small or medium model; test accuracy on your own audio first. Regular meeting transcription with a large or turbo model is better on a GPU. Every ANVE tier, as specified for launch, includes a GPU, either the integrated graphics of a Strix Halo-class chip or an NVIDIA card, and ANVE has not published transcription timings for its launch hardware.
05
Questions people ask
Is a CPU-only machine good enough for a small office?
For a few recordings a day with the small model, the benchmark suggests it can be, since the slowest setting still finished faster than the recording. For large-model accuracy, Hindi or many recordings, add a GPU.
How can I speed up Whisper on a CPU?
Use faster-whisper with 8-bit precision and batching. In the benchmark that took the small model from 6 min 58 s to 51 s for the same 13 minutes of audio on the same CPU.
Does a bigger Whisper model run on a CPU at all?
Yes, but more slowly. OpenAI’s table gives large a relative speed of 1 against about 4 for small and about 8 for turbo, measured on an A100; the benchmark above did not test large on a CPU, so no figure is quoted here.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
