A major downside of vanilla SDR is that it’s sensitive to the loudness of the predicted signal.

Hi Evan Radkoff, don't you think, in general, source separation tasks require preserving the loudness.

At this point, I see no reason why anyone would want to use plain SDR in favor of the scale-invariant version.

In my opinion, during learning SI-SDR could be counter intuitive, but as an evaluation metric, it makes more sense for SDR to be scale invariant.

Hi Dan. For evaluation, I suppose it depends on the end use case. I could see scenarios where scale is irrelevant. For example if you're feeding the separated signals into a classifier, you can just normalize as a post-processing step. On the other hand if you're trying to measure the loudness of each source, obviously SDR is more appropriate for choosing the right model.

For learning I think the only real answer is "try both and see"