Originally Posted by Bass Thumper
Originally Posted by Cerio
Originally Posted by DrDan
Quote

I wouldn't base the quality of a stem separator based on the quality of just one instrument on one track.

Wise words...
I now have a 2nd datapoint.
I extracted my own bass line from my most recent composition Down Periscope

Just to clarify, I'm not saying you're wrong, maybe you're right (and Presonus algorithm is the best one currently availiable for bass), maybe not, I really have no idea.
What I want to point out is that drawing such a general conclusion from such a particular case makes little or no sense to me. If someone really wants to compare the different algorithms available on the market to figure out which oneis ‘the best’ for bass, then it would be necessary to compare not one or two examples, but a large number of them, with different audio qualities, different mix qualities, from different eras, different compression ratios, different styles of playing, different bass sounds and techniques, different panning, etc. And of course, if this job is going to be made by different people, the test files must be identical for all of them; posting a youtube link and expecting that everybody else will download the same mp3 with the same quality from that link is probably not a realistic idea.

Bearing in mind that (as far as I know) the vast majority of available algorithms actually use the same common code (derived from the Open Source Spleeter -or, in some cases Demucs-) and differ from each other basically in terms of the amount and quality of material they have been trained with, the result of such an experiment would probably be that there is no ‘best’ algorithm in a broad sense. Some of them would simply work better in some cases than others, and viceversa.

It seems common sense to me that an AI trained mainly with classic jazz material will do a better job separating Paul Chambers' double bass line ion "So What" than another AI trained mainly with 70s funk. I think that the next generation of stem separation programs will allow the user to choose between different trained models, depending on the starting material.

Last edited by Cerio; 10/13/24 12:42 PM.
Latest BIAB version, latest build.