Let's continue the #discussion but with a few new clarifications...

Can a self-supervised neural representation that maps audio/spectrograms to fixed dim embeddings at a fixed temporal resolution *simultaneously* solve a range of existing audio tasks (e.g the HEAR Benchmark)?

If so, why hasn't this worked yet? If not, what else do we need?