Frequently asked questions

Why don't you have autotags for my favorite tag?

We can only train an autotagger if players of the game have identified enough examples of it. "Enough" generally means 30-50 clips that have been verified by two people as relevant to a tag. So if your favorite tag isn't up there yet, get playing and find some good examples of it.

If I play the game now, will my tags get incorporated into the search?

Yes they will. Well, they will be incorporated into the human search within a day. They won't be incorporated into the machine search until we re-train the autotaggers, which only happens occasionally.

Did you train and test on the same data?

No, all autotagging results are shown for a separate set of artists than they were trained on. Assigning artists to either the training or testing set makes sure we're really learning the sound of the tag and not something like the "producer effect".

How do the autotaggers work?

First, we extract timbral and rhythmic features for each of the clips. Then we train support vector machines (SVMs) on these feature vectors using the data from the game. You can read about it in our forthcoming paper in the Journal of New Music Research.

What order are the autotags in?

They're sorted from the most accurate to the least. This was determined by measuring classification accuracy on labeled clips in the test set.