Key Analysis Accuracy Comparison 2019

General Mixxx discussion.... discuss

Moderator: garth

Key Analysis Accuracy Comparison 2019

Postby bascurtiz » Sun Jul 21, 2019 11:40 pm

Hi Mixxx Development Team!
I've included Mixxx in my Key Analysis Accuracy Comparison 2019 (12 apps compared with 840 tracks) too!
Check it out here: https://www.reddit.com/r/DJs/comments/c ... ison_2019/
bascurtiz
 
Posts: 3
Joined: Sun Jul 21, 2019 11:38 pm

Re: Key Analysis Accuracy Comparison 2019

Postby tapir » Mon Jul 22, 2019 6:38 am

Thanks for the info and your work! We are already aware of your survey and discussing the results on our Zulip channel.

I checked just a few tracks with 2.2.1 and found at least one track that gives different results. We have no explanation for this deviation yet.

Hopefully we could repeat this comparison with 2.3.0. We recently contributed some improvements to the QueenMary key analysis plugin before any knowledge of your comparison ;) The results are promising.

PS: You got an upvote from me on Reddit
tapir
Mixxx Developer
 
Posts: 325
Joined: Sat Aug 04, 2012 12:03 pm
Location: Stuttgart

Re: Key Analysis Accuracy Comparison 2019

Postby bascurtiz » Mon Jul 22, 2019 8:45 am

Awesome!

...and that's the spirit =)
Great to see you guys aiming for better results.

Strange you had certain different outcomes result-wise.
Could have to do with the bitrate of yours vs. mine.
If you want, I can provide the tracks used in this test, so you have something to test in-house with.
bascurtiz
 
Posts: 3
Joined: Sun Jul 21, 2019 11:38 pm

Re: Key Analysis Accuracy Comparison 2019

Postby daschuer » Wed Jul 24, 2019 3:10 pm

I was the guy commiting the changes to QM and I have some questions:

How was the manual key detection performt?

Especially distinguishing minor from mayor tracks is a real challenge. The algorithm is able to detect chords, but often the musicbrains do not follow pure minor or minor progressions. Sometimes you even find a chorus in major and the other parts im minor of a compatible key like C and A min. Is such a track minor or mayor?

Luckily this does not really matter because C and A minor at compatible in the same way to any successor track.

What do you do you treat gypsy minor or harmonic minor tracks not fitting into the circle of fifth?

I am tempted to leave them undetected, because treating them as one of the circle of fifth keys leads to a wrong asumption about their compatibility.

Thank you for you help.

Daniel.
daschuer
Mixxx Developer
 
Posts: 1226
Joined: Sun Mar 10, 2013 11:26 am

Re: Key Analysis Accuracy Comparison 2019

Postby bascurtiz » Wed Jul 24, 2019 8:32 pm

Hey Daniel,

How was the manual key detection performt?

I've never been part of that. I just used the data they put together for Ibrahim's thesis, see: http://ibrahimshaath.co.uk/keyfinder/KeyFinder.pdf and http://ibrahimshaath.co.uk/keyfinder/Ke ... ataset.pdf

...like C and A min. Is such a track minor or mayor?

If you mean C Major vs. A minor, then yes, this is indeed a tricky one... Hence I counted the relative Major/minor that were wrong in the 12 apps test again, and counted them as right, for the follow-up test, Top-5.
The difference of correct determined key is therefore 4% max. more (based on the Top-5).

What do you do you treat gypsy minor or harmonic minor tracks not fitting into the circle of fifth?

Looks like any program out there only restrict the determination to either Aeolian or Ionian scale.
I think the people determing these keys, therefore tried to do so aswell.

I am tempted to leave them undetected, because treating them as one of the circle of fifth keys leads to a wrong asumption about their compatibility.

I think that's the way to go indeed.
Mind you, some apps didn't detect a key (perhaps for this reason u mentioned for example).
I didn't count them wrong nor right. I just subtracted them from the total count for that specific app.

Thank you for you help.

You're most welcome!
bascurtiz
 
Posts: 3
Joined: Sun Jul 21, 2019 11:38 pm

Re: Key Analysis Accuracy Comparison 2019

Postby daschuer » Fri Aug 23, 2019 6:09 am

What does this mean to you.

All programs pick the most likely major and minor scale.
If the track is non of this, the result is random and useless. Since you don't see a difference between a proper detected minor track and a random pick for an undetectable track, all results become unreliable.

I like to fix it for Mixxx. What would this mean to you?
daschuer
Mixxx Developer
 
Posts: 1226
Joined: Sun Mar 10, 2013 11:26 am


Return to General Discussion

Who is online

Users browsing this forum: No registered users and 2 guests