about sms_wsj about asteroid HOT 3 CLOSED

When creating this database, we did not consider groups who only have the rights to wsj0 but not to wsj1.
One of our criticism regarding other source separation databases was that they do not allow for acoustic model training because of small training set and low variability.
Using both databases gives us a larger number of speaker and utterances to chose from and allows for a large enough training dataset to train a strong acoustic model.

For now there is no possibility to work on this database without WSJ0 or WSJ1 since otherwise the comparison between the results would no longer be possible.

from asteroid.

It's mentionned in the abstract of the paper ("WSJ0+1").
You can't generate SMS_WSJ without WSJ1.
Can I suggest that you have a look at Kinect-WSJ which is based on WSJ0 and CHIME5, also multichannel with speech mixtures in noise.

cc @jensheit who is one of the authors, and the one who implemented the recipe in Asteroid.

from asteroid.

thx

from asteroid.