Live session L12 · day 57
Audio fingerprinting from first principles, tested against a loud room with someone talking over it.
Releases Wednesday, 25 November 2026
Jamz Finder identifies a song from a phone held up in a noisy room, and it does it with no model and no embeddings. Constellation hashing from a spectrogram, which is a 2003 technique and still the right answer.
Phone to a speaker, across a loud room, someone talking over it. Every session breaks something on purpose. A build that only ever works is a demo, and a demo is the one thing this is not.
The recording is posted on this page, and the write-up goes out the same day at designsbyduhart.org/blog.