Concepts
Privacy and performance
What stays on your Mac, how fast transcription is, and how Stenny keeps your Mac usable.
Privacy
- Audio never leaves your Mac. Whisper large-v3-turbo runs locally on the graphics chip, through mlx-whisper.
- The AI step sends text, not audio. If you turn it on, the transcript text and the slide text go to Claude. With
backend = "none", nothing leaves your Mac. - Your Claude project reads from your vault through the Filesystem connector, only when you ask it something.
Recording a lecture may need the lecturer's consent. Many universities, WU Vienna for example, require it. Ask first, and keep recordings and transcripts to yourself.
Speed
Measured on a MacBook Air M4 with 16 GB, on battery. A three-hour lecture takes about 20 to 25 minutes at the default priority.
| Priority | How to use it | Speed | Effect |
|---|---|---|---|
| normal | --now | 9.4× real time | Full speed. |
| low | The default | 8.3× real time | Other apps get priority. |
| background | Queued jobs | 7.3× real time | Processor, disk and graphics chip throttled. |
Resources while recording
| Resource | Use |
|---|---|
| Memory | About 2.5 GB for the Whisper model. |
| Processor | On average about 6 % of one core. The backup microphone adds about 1 %. |
| Graphics chip | Short bursts for each chunk of the live draft. |
| Disk | About 12 to 14 MB of audio per hour, about 1 GB per semester. |
To save battery during a lecture, untick Live Transcript While Recording. Stenny then only records and transcribes after the lecture. To run queued jobs only at night, set night_only = true.
Accuracy
The benchmark uses 18 minutes of real French speech, Maupassant's La Parure from LibriVox, compared with the book text. All engines use the same large-v3-turbo model.
| Engine | Speed | Peak memory | Word error rate |
|---|---|---|---|
| Stenny (mlx-whisper, chunked, with glossary) | 9.4× | 3.4 GB | 2.7 % |
| mlx-whisper, long-form | 7.9× | 3.4 GB | 3.1 % |
| whisper.cpp (Metal) | 6.1× | 2.2 GB | 3.2 % |
| faster-whisper (CPU) | 2.0× | 3.3 GB | 2.9 % |
Most of the remaining errors are number formatting, like “3” instead of “trois”, and French agreement endings.
German and English
The same pipeline on two more LibriVox recordings, each compared with its Project Gutenberg text. No glossary, as there are no course terms in fiction.
| Language | Recording | Length | Word error rate |
|---|---|---|---|
| German | Kafka, Ein Landarzt | 18 min | 3.1 % |
| English | O. Henry, The Gift of the Magi | 13 min | 3.1 % |