2 ms·
Gotcha. So taking n overlapping samples for the song, fingerprinting then vectorizing each. Then doing the same thing for the captured sample and getting neare
by bequanna 3y ago
Gotcha.
So taking n overlapping samples for the song, fingerprinting then vectorizing each. Then doing the same thing for the captured sample and getting nearest neighbors.
I’ve used dejavu for a personal project and it is quite fast. Is the LSH approach somehow more efficient/attractive now because we’ve made strides in vector nearest neighbors search?
- willseth 3y agoAlmost, but dejavu doesn't search nearest-neighbors. It searches for exact matches, so its robustness is essentially based on oversampling. Nearest neighbor search might be a way to achieve similar robustness with fewer fingerprints - assuming it results in fewer false negatives. LSH is one way to approach nearest neighbor search. So in theory you could modify dejavu with an LSH-based NN search and potentially increase robustness with less data, modulo whatever tradeoffs are necessary for computing NN search.