beaTunes 3 Key Detection Accuracy
I promised yesterday that I'm going to back up my claim that beaTunes 3's key detection works pretty well. To make it short, I tested against the Beatles studio albums. They make testing easy, because reference keys are published at Isopohonics. It also means that you can easily verify my results.
The graph shows correctly recognized keys.
Not so bad, right?
Does it mean the key detection of beaTunes is significant better than the key detection of MIK?
More or less.. It means that for the Beatles test set, beaTunes 3 with online resources turned off performs a lot better than MIK 4.1. Things may look different with other test sets.
That said... I'm actually surprised none of the DJ sites/mags have picked this up yet. Somebody should tell them.. ;-)
Made small comparison between version 2 & 3: http://i54.tinypic.com/2rz50jp.png - difference is noticeable. Waiting for full comparison before buying v3.
Are you planning to conduct a full comparison? I will not..
No, I will wait for someone qualified :)
Seriously, the comparison with b2 will be useless unless you also compare against a reference dataset (i.e. a ground truth). Only then you can make meaningful statements about accuracy of either version. That said, in internal testing b2 performed pretty much as well as MIK41 on the Beatles testset, i.e. roughly 53%, if I remember correctly.
I don't see error bars... what's your p value there? :o)
Stevie, besides this blog post not being a scientific paper, I'm not sure you understand the nature of this graph and the experiment it describes.
The experiment is repeatable, i.e. it always delivers the same results. And the actual probability of any of the algorithms delivering the correct result (i.e. the null hypothesis) is unknown.
It appears to me that a p-value therefore does not make any sense.
Hi is there a way to, instead of having beatunes detect key in letters, have it detect it in numbers such as 8b or 9A, etc.
beaTunes detects keys, not letters, not numbers.
It then displays this information in their proper musical name and also in the Open Key Notation, a number based format.
beaTunes does not display the key information in Camelot notation, because the inventor of that notation requires other companies to license that particular notation. Feel free to google "beaTunes Camelot notation" to find more information about this topic.
So how does it's detection rate compare with dance music?
I wish I could tell, but I'm not aware of a publicly available reference set. And as long as that's not available, anybody who claims to do really well on dance music is spreading myths.
That said, if you know of a good reference set, I'd be more than happy to run the test and post the numbers.
Could you please run a comparison against MIK 5? I own and love your software, but I am very curious to see if MIK 5 has managed to close the gap in terms of accuracy.
Hi, I compose through MIDI devices and softwares. Can your program determine the key of a MIDI file ? Or is it audio only ?
It seems your program doesnt work that well (regarding key detection, I still love it to keep my metadata consistent and the creative suggestions for playlists):
The test contains also a list of the 50 songs which are used for testing.
Yep, beaTunes does not do well in that test. However, with just 50 tested songs, the 95% confidence interval is probably around plus/minus 12%. (please correct my math, if this sounds off to you)
Songs on beatport have the key listed? Could you use that as a test?
" Steve C said...
Songs on beatport have the key listed? Could you use that as a test?"
I would be interested in this as well :D
Beatport uses software to detect the key. It's not provided by the artist.
Any idea which software they use?
(given the state of the art - it could also be wrong)
Post a Comment