sigsep / sigsep-mus-cutlist-generator Goto Github PK

@faroit The times in the 7s preview cut-list are relative to the start of the 30s previews. Should we make this clear in the README?

I overlooked this when extracting the BSS Eval scores for the 7s previews from the main results file.

Hey,

In a fresh conda environment, pip install -r requirements.txt failed.

It seems I need to install llvmlite first, e.g.
conda install -c numba llvmlite

the activation code was submitted as PR for medleydb
It seems that there are some modfications done since then, so it makes sense to update to the final version in medleydb

to match with bsseval segements, the 30s excerpts should be rounded to full seconds thus add some code here

I did some rehearsal of the selections and found some significant flaws with the 7s preview segments.

e.g.:

in PR - Happy Daze the output selects: start: 95; stop:125sfor the 30s segments and [22s; 29s] for the 7s previews. The problems is that the 7s preview part does not contain any vocals. So I assume there is some bug somewhere. As far as I checked, the 30s previews are okay.

@deeuu any idea?

sometimes its nice to directly use the csv from a url (as pandas can read from url). So I would propose to use the plain text csv instead of zip.

The output should be based on the 30s excerpts, and generates a 2nd preview dataset

I think for some applications we might also want the training previews for 30s and 7s duration.

@deeuu do you have computed them as well and can maybe update the release? 🙏

@faroit Should we allow for amplitude envelope ramps to be applied to the onset and offset of the cropped audio?

I'm happy to leave this as a post-processing step for the listening test.

Should we call this one:

sigsep-mus-2018-cutlist-generator

sigsep-mus-2018-trim should then probably be renamed to sigsep-mus-2018-preview-generator (which you might want to fork to the sigsep organisation).

If you're not fussed then let's leave it!