kepstin: I must say, finding differences in those songs is properly hard, even after having the exact times from the matching algorithm :)
a7medo778 has quit
kepstin
heh, and you don't want to make the matching oversensitive, since that'll make it throw away matches for things that are actually the same :)
luks
and the new matching algorithm will depend on <5s segments that are distinct enough to tell those songs apart, so it won't be very robust :)
I'm not working with averages over the entire fingerprint this time, so it shouldn't be a problem, I basically split the fingerprints into segments and if there are some segments that do not match well, it will not consider those two fingerprints to form the same "acoustid"
Freso
luks: Will there be some kind of versioning so it's possible to tell when generation an AcoustID is?
luks
probably not, this won't make it to acoustid for a long time and when it does, it will only affect a very small chunk of the database
reosarevok
Clint: paella is from Valencia! :p
kepstin
so, it's sort of like it'll do the average matching to find match candidates, then if it's a case where some small segments are very different and the rest is the same, it'll throw away the match?
luks
well, it will still give you the match when searching, but it will be more strict about merging individual fingerprints into "acoustids"
kepstin
ah, so since it'll keep them separate, you'll get a bunch of acoustids with different match% returned, instead of only one acoustid back?
luks
you already get multiple acoustids back with some score
kepstin can't really remember the lookup api well, he should take a look at the docs again :)
so everything more or less stays the same, except that the internal search when processing a new submission will be more strict
on the second graph, the blue line is an error value for the match, I guess?
luks
it's the difference between two levels of "blur" (the blue and green lines in the plot above), it's normally used in CV for blob detection
and it's smoother out a bit, to avoid noise around 0
smoothed
the reason I'm doing this is so that I can tell when two fingerprint start matching, so I can locate a short sample in a long one and stuff like this
telling those songs apart is just an added bonus :)
kepstin
cool
kepstin manually looks up a 'the world is all one' fingerprint out of curiosity - it returns one result only, with a list of multiple musicbrainz recording ids.
using my awesome application #39, "Kepstin is using curl to manually do lookups on the command line"
Leftmost
Is that available on github? Or do you run it as a service? :-P
kepstin
it's currently version -2, which I think means you have to use a time machine to obtain a distributed copy.
kartikgupta0909 joined the channel
krono has quit
krono joined the channel
Clint
reosarevok: i agree
Gilou
are there any protection in abz / acoustic id to avoid clearly wrong music from a bad mbid ? (bad as in, not set to what the song really is)
not sure how that could work.. but I have seen some weird values at some points, and I assumed it was because something went wrong there, rather than the algorithm being totally fooled.. (I don't have the example handy, hence my asking around a little bit vaguely)
Freso
Gilou: No.
Gilou: For AcoustID, you can log in on acoustid.org and de-associate wrong matches.