Audio Source Separation

Hi all,

I would like to propose a new task, actually more like a "sub-task" for the professionally produced music recordings. The thing is that many works on source separation for music are actually background music/foreground voice separation. Such a task is often sufficient for applications such as melody extraction (for the voice) or karaoke gaming (for the music). Since there is a fair amount of researchers working on that, I think it could be a good thing to have a task (or sub-task) specifically dedicated to background music/foreground voice separation.

To begin with, a simple additional column could be for example added to the results that measures the separation performance for the music overall (in addition to the individual instruments). More data would be obviously more than welcome.

Thanks.

Zafar Rafii

Hello Zafar,

The task that you propose is inherent in the currently posed task, since if you know to extract the voice from the mixture, the background music is given, too. How? Well, subtract the voice from the mix and you have it. This is of course only true if the mixture is a simple summation of the distinct tracks (which is the case here, I think).

Best,
Stanislaw

I meant I would like for researchers to also be able to report their results for the full music background on the website. As far as I remember, if there were more than one instrument for the music background, researchers doing music/voice separation could only report their results for the voice component.

Thanks.

I would like to see PEMO-Q's PSMt metric included in the evaluation. It is the most reliable metric that I have tested so far. I certainly prefer it over the PEASS toolkit.