Spent the morning playing with cosine similarity searches over audio embedding: feed a track, get a list of similar samples. Currently running Drop The Hate against tidal-drum-machines (default for strudel)

% time python query_vibe.py fbs_drop_the_hate.wav --collection tidal-drum-machines --top-k 10
Loading CLAP model (laion/clap-htsat-unfused)...
Loading weights: 100%|█| 447/447 [00:00<00:00, 10124.11it/s, Materializing param
Loading fbs_drop_the_hate.wav...
  Duration: 330.4s
Querying collection 'tidal-drum-machines' (2595 samples)...

Top 10 matches for: fbs_drop_the_hate.wav

Rank  Similarity  Sample
------------------------------------------------------------
1       0.2778    machines/YamahaRM50/yamaharm50-lt/TOMS_103.wav
2       0.2772    machines/SoundmastersR88/soundmastersr88-oh/Open Hat.wav
3       0.2657    machines/YamahaTG33/yamahatg33-fx/SFX-01.wav
4       0.2633    machines/RolandMC202/rolandmc202-bd/Bassdrum-04.wav
5       0.2503    machines/RolandMC202/rolandmc202-bd/Bassdrum-01.wav
6       0.2486    machines/YamahaRM50/yamaharm50-lt/TOMS_104.wav
7       0.2448    machines/DoepferMS404/doepferms404-bd/Bassdrum Reverse.wav
8       0.2442    machines/RolandCompurhythm78/rolandcompurhythm78-misc/Quid-02.wav
9       0.2429    machines/RolandMC202/rolandmc202-bd/Bassdrum-05.wav
10      0.2428    machines/RhodesPolaris/rhodespolaris-bd/Bassdrum-01.wav

4.27s user 0.64s system 52% cpu 9.309 total

not really good yet, but it's amazing how amazingly vector databases amaze me 😂