#musicbrainz

The offset part is especially cool.

http://acoustid.org/fingerprint/20186887/compar... hmm.

kepstin-work: maybe the really really close parts compensate for the not-so-close parts

i.e. the error rate is averaged over the whole file and manages to be low enough

yeah, I suppose so.

kepstin-work: That one of the problematic karaoke tracks? ;-)

yeah

well, that's my horrible different vocalist case, I think

I don't see what's wrong with it, but I'm not very good at reading acoustids.

Leftmost: different vocalists to the same backing audio, I guess.

This one is really weird.

Move up to -12, then -24, then -36…

-48 too!

and -60

yeah, it’s consistent.

whoa.

how does that even

i wonder if it's a software bug somewhere

dunno. I had another one where it was clearly stretched between the two. You could see the dark area shifting as you adjusted the offset.

But nothing like that 'perfect match, shift, perfect match, shift' thing.

i'd expect stretched to be fairly common

particularly for older stuff that went through an A2D conversion

Here it is: 'move up' — http://acoustid.org/fingerprint/10446680/compar...

yeah, that's consistent with a resample at slightly different rate or slightly different playback speed on analogue equipment

Another one with that other weird offset, this time at 33 intervals (move up, start at -2)

the only way to explain that is that chunks of the audio have been deleted/cut out from one of the files

or a bug in the audio decoder in fpcalc :/

"acoustid submissions would link the fingerprints to musicbrainz track ids" +1

And with that, bed time.

make that +2

repetitive parts in a song are not that unusual

luks: It’s not just repetition, or we’d see larger and overlapping chunks wouldn’t we?

in this, we're seeing different chunks of the song matching at different offsets

right.

i.e. 0-10 matches 2-12, then 10-20 matches 28-38 or so

so one files looks similar to the other file, but with regularly spaced chunks of missing data

hmm. the 'differences' view in the fingerprint comparison is equivalent to doing an xor operation between the two, isn't it?

I think it’s an AND

no, an AND would make it white when both sides are the same :)

depends if black is 1 or 0

if you want 0,0->0 0,1->1 1,0->1 1,1->0 then you want xor :)

and and or both miss a case.

Ah, so they do.

what do you know, fpDiff[i] = fp1[i + offset2] ^ fp2[i + offset1];

xor it is :)

:-D

it is.

hawke_1: I think these kind of matches are cause by e.g. radio edits

if you compare original edit and radio mix, all of them would match at some point

luks: that doesn't create a repeating pattern of different chunks matching at different offsets.

it would also be weird to have a radio edit of a cat stevens song. ;-)

why would you remove the first 2 seconds, then the 8-10 seconds, then the 18-20 seconds, etc?

I think it does, but if I'm not even able to type, I might be able to think right :)

kepstin-work: well, that's what I would call audio editing

if you need to create a short version under 3 minutes

for 70s stuff, wasn't there often a radio or "single" edit to keep the song within 2-3 minutes?

right, but they'd usually do stuff like remove an entire verse for that, not remove e.g. 2 seconds of ever 10 in a repeating pattern :)

you never know

I used to debug a case when the person who sent me two versions and asked why chromaprint thinks they are different couldn't even tell they are actually different

even I couldn't tell it by listening, I had to play them side by side in audacity

... and the difference was?

lol

kepstin-work, luks: http://imagebin.org/273136

so strange.

Fingerprint only submitted once.

maybe they ripped it from ascratched cd :)

Hmm, would that do that?

Seems strangely consistent.

The recording is http://beta.musicbrainz.org/recording/3dc50f9e-... — if the acoustID duration is correct, that means it’s either “Classics, volume 24”, or “Collected”

luks: Do you store the release MBID with the submission, by chance?

It appears that it’s not “Classics, Volume 24”

I understand that not everybody who scans to pdf will take the time to make the page numbers agree. But if you're the Getty and the document your scanning is a about a schema for art metadata? I hold you people to a different standard.

grr :s/your/you're

What’s a Getty?

i assume the reference is to the museum

http://getty.edu/ is a getty.

but there must be others

haha. -1 to hawke's rfc

mb-style is on a roll

-1 to that. Silly. Hopefully she won’t veto it.

god forbid we improve things incrementally.