Anyone there?

The register

Whisper base listened to 504 windows of nothing: digital silence, and white, pink and brown noise and mains hum at different levels, eight seconds each, in 24 languages. Not one of them had a voice in it. It said something in every one.

60would reach a transcript with Whisper's default settings
97Whisper would decode again, hotter
347its guard would drop as silence
62%of the time, the static itself sounded like Norwegian Nynorsk to it

A name from the creditss

Subtitles for TV end with who made them, over the silent last seconds of a show. Whisper learned that silence sounds like those names. These are the spirits it believes in most: it is sure nobody is talking, and just as sure about the names.

The end of a videos

Silence usually comes right after someone says goodbye to the camera, so silence brings the goodbye back.

A sentence nobody saids

No credit, no goodbye. Just the likeliest words for a room with nobody in it.

A chants

The decoder falls into a loop and can't get out. Whisper's guard calls this too repetitive and tries again.

A numbers

Plain hiss, heard as Norwegian Nynorsk, comes out as a number.

Run it yourself

Same model files, same static, same spirits. npm run sweep rebuilds this page's data from scratch on your own machine in about six minutes, and the tests check that every entry here really turned up. The model is whisper-base, quantised to 8 bits.

Back to the séance