Timbre

Audio similarity search over 105,884 Free Music Archive tracks, on a hand-written HNSW index.

or upload a clip (only the first 30 seconds are used):

What just happened. Your phrase was embedded with CLAP into the same 512-dimensional space as the music. The index holds 2,143,867 overlapping 10-second segments — about 20 per track — and a hand-written HNSW graph found the nearest ones, then kept each track's single best-matching passage. The timestamp on each result is where in that track the match was.

No audio is hosted here. The corpus is 99 GB, so results link to a YouTube search for the artist and title. The search may not find the exact recording.