darius522 / dnr-utils Goto Github PK

Hi, @darius522 congrats on the repo and corpus, is really useful!

We got a lil question regarding some mixtures of the corpus. We were arbitrarily listening to some test samples, and we found that in some of them when they have music as the background, it's not only instrumented but they also have singing.

Hence, the mixture has, say, 2 'speeches' one from the speech.wav utterance and another one contained in the music.wav track.

Such mixtures are, for instance, 3030 - 11623 - 1882 - 9963.

We intend to experiment with DnR on two separate tasks: speech separation and source separation.
One doubt we have is, having the types of mixtures described above, where an extra kind of 'speech' is present in the mixture (e.g., the one belonging to the background music), isn't this a potential issue for speech separation algorithms?

Likewise, is this a potential issue at all for source separation algos?

We were just curious about this, hope you can provide us with a guide to a better understanding. Thanks in advance! :)

Hi,

The link and corresponding download command in the README (https://github.com/darius522/dnr-utils#download) points to V1 version of the dataset.

I used that command to download the old version, and now I have to re-download V2.
Or is there a way to just download the difference between V1 and V2?

Also, is V1 a subset of V2 or are they completely different?
Since the number of extra files is not too large, I am thinking if I should just use V1.

Some people may have trouble training models on such massive amounts of data.