Prompt & Done · STENO pack

An hour of audio, transcribed locally in 28 seconds.

OpenAI's Whisper (large-v3-turbo, MIT) running on Apple silicon through the free mlx-whisper package. Tested on an M5 Max with 64 GB, 26 Oct 2026; speed depends on your Mac.

The command

brew install ffmpeg
pip install mlx-whisper
mlx_whisper your-audio.mp3 --model mlx-community/whisper-large-v3-turbo --language en --output-format txt

The first run downloads the model weights. Or download transcribe.sh. Docs: ml-explore/mlx-examples, whisper.

What we measured

TestResult
AudioLibriVox, Alice's Adventures in Wonderland, chapters 1-4, public domain, 60 min 33 s
Wall time (whole CLI run)28.9 s, 28.3 s, 28.2 s: about 128x real time
Word error rate vs the Gutenberg text4.2% (some of it is LibriVox announcements and punctuation style)
Claude Haiku summary of the transcript5.6 s, about $0.007 API-equivalent

The full transcript is here: alice-ch1-4.txt.

Honest limits

One clear English reader. Accents, noise and overlapping speakers raise the error rate. Not tested on other Macs or on non-English audio.