Entra:     

Forum: Wishes and new features

Topic: Stems feature suggestion
Can I suggest a future update to stems? The ability to remove only the lead vocals, but leave in the background vocals. This would be a HUGE help for karaoke singers, and those who want to sing to the original tracks. Thank you!
 
Inviato 7 hours ago
It would mean to train a new model for stems in order to do it.
Training on which dataset (as it's supervised training). Not sure there is a good one with enough examples. Even if you only work on the vocal part.
 
I think that would be very difficult to achieve because often the same vocalist performs lead and background.
 
That's the beauty of machine learning. Maybe it can find a model. But for that we need the following list with many samples :

1 | vocal1.mp3 | vocal1_lead.mp3 | vocal1_background.mp3 | vocal1_length
2 | vocal2.mp3 | vocal2_lead.mp3 | vocal2_background.mp3 | vocal2_length
...

Where can we find it?
 
A good starting point is producer stems - they normally separate lead and background vocals, and are clean.

You'd need A LOT of those though (lol good luck there).
 
Maybe here after a first search:
 
Also, would you want to run that as a new model analysis behind the original Stems 2.0 model (focusing only on multi-vocal separation), or update the Stems 2.0 model itself (as an update to the vocal separation done there)?

Former might be faster to do but have more time upfront (might have to wait for the first to finish or some significant part of it to finish).
 
Would be great to have an advanced stems separation for those with decent GPUs. It will inevitably happen at some point but perfectly happy with what has been achieved so far. Especially compared to other DJ software. After using Denon standalone for a while it makes you realise how brilliant the VDJ stems really are in comparison.

I use them pretty much every gig and it really is totally seamless.