Why reconstruct linear spectrogram? about s3prl

[andi611]

The initial thought is, we hope that the model can capture more information through reconstructing a more frontend acoustic representation.
However, in our later experiments, we find that Mel->Mel and Mel->Linear yield similar performance for phone classification (This Exp is not included in the paper).
So for now I would say Mel->Mel is sufficient (The Mockingjay BASE is Mel->Mel and it also performs well).
I'm not one of the authors of Audio ALBERT, but I believe they simply follow the implementation of Mockingjay, and there is no particular reason for choosing such a setting.

[iamjack2018]

That makes sense. Thank you for your clarification.