Guides

Run Whisper on your computer

Pick one. All three are free, open source, and run with no internet once the model is downloaded.

No app? On a Mac, MacWhisper and Aiko wrap Whisper in a normal window and say they transcribe on the device. The rest of this page is for the terminal. Sizes are on the Whisper page.

whisper.cpp

The lightest option, from the same people as llama.cpp. Runs on CPU, Apple Silicon (Metal and Core ML), and NVIDIA. Official repo: github.com/ggml-org/whisper.cpp. The quick start there, copied on 19 Nov 2026:

git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
sh ./models/download-ggml-model.sh base.en
cmake -B build
cmake --build build -j --config Release
./build/bin/whisper-cli -f samples/jfk.wav

whisper-cli wants 16-bit WAV. The README converts other audio with ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav. Swap base.en for large-v3-turbo when you want accuracy. There is also a whisper-server example if you want a local HTTP endpoint.

faster-whisper

Python, built on CTranslate2. MIT licensed. Its README says it is “up to 4 times faster than openai/whisper for the same accuracy while using less memory.” That line is Vendor-reported. Repo: github.com/SYSTRAN/faster-whisper. Needs Python 3.9 or newer. NVIDIA GPU use needs CUDA 12 and cuDNN 9; the CPU works without them.

pip install faster-whisper
from faster_whisper import WhisperModel
model = WhisperModel("turbo", device="cpu", compute_type="int8")
segments, info = model.transcribe("audio.mp3")
for s in segments:
    print(s.text)

The README shows "turbo" on CUDA and device="cpu", compute_type="int8" for CPU. I combined the two lines. That combination is Not yet measured here.

OpenAI’s own package

The reference version. Python and PyTorch, plus ffmpeg. Heaviest of the three: OpenAI’s table lists about 6 GB of VRAM for turbo and about 10 GB for large. Vendor-reported. Repo: github.com/openai/whisper.

pip install -U openai-whisper
whisper audio.mp3 --model turbo

Ollama and LM Studio

Ollama does not have an official Whisper model. ollama.com/library/whisper returned “not found” on 19 Nov 2026. A few user uploads use the name; I am not linking those. Use whisper.cpp or faster-whisper, then hand the text to your chat model.

On Frankenstein

Measured on Frankenstein (HP ProLiant DL380p Gen8), 19 Nov 2026. This is not the quick start above. The run used commit ca19af24, CPU only, -t 24 -l en, median of 2 runs, on a 161.2 s public-domain LibriVox reading of the Gettysburg Address. The build line and the exact command are on the methodology page.

On a CPU this old (no AVX2), medium beat turbo. The Tesla P40 run is Not yet measured. faster-whisper and OpenAI’s package are Not yet measured too.