Question about the drum stem/dataset we are submitting to

When i listen to the MUSDB18 drum stems, sometimes i hear the instrument shaker. Please someone can tell me if for the dataset we test against, shaker is also on the drums stem?

1 Like

Also i extend this question for wind chimes.

I run this file through demucs and demucs output the wind chimes onto drum stems.
Demucs output

I would believe this wind chimes belongs on the other stem. Someone can clarify what instruments belongs to this drums stem? As we do not want to train our model with incorrect data affecting our performance to the dataset we test against.

I think it is usually up to the original artist as to which stem they put which instruments into. Some might think of wind chimes and shakers as percussive (since they are not persistent like the melody) and hence club them with the drum stem. I have also seen cases in other stem files, where some artists will put kick/snare sounds in vocals stem, just because they created the kick/snare sounds by filtering/modifying a vocal sample.

So I guess we should expect some bleeding among the stems to happen. This is one of the focuses of this year’s competition, they have separate tracks for bleeding and wrong labels.

However, I guess in the MUSDB18 stems this problem is not too significant, as for the test set, it might have some degree of similar instrument bleeding.

I am also just a participant and this is just my opinion.

1 Like

Hi @lyndon_reid and @subatomicseer Thank you for your question.

Shakers and wind chimes are on the ‘drums’ stem in musdb18.

Percussive instruments such as shakers and wind chimes usually belong to the label ‘drums’ in musdb18. For example, you can check ‘shakers’ in the drum track ‘Invisible Familiars - Disturbing Wildlife’ from musdb18.

However, pitched percussion instruments such as marimba, chime bells, and Xylophone belong to the ‘other’ label. For example, you can check ‘Xylophone’ in the other track ‘The So So Glos - Emergency’ from musdb18.

The dataset for this challenge probably might follow the same convention. However the organizers are checking details currently. We will keep you posted when we are ready.

As @subatomicseer mentioned, I also think the problem is not too significant in musdb18. Especially for the instrument-stem matching protocol, it would not be problematic as long as the protocol is consistent in both training and test dataset.

The dataset of this challenge (before data corruption) will be created under the same protocols.

1 Like