WEIRD THOUGHT
A finding of machine learning is that every recognizer is also a generator. You can convert an algorithm that recognizes giraffes into an algorithm that creates pictures of giraffes.
Psychoacoustic audio codecs, like MP3 or AAC, are in some sense recognizers.
If you treated MP3/AAC like a statistical model— generated random data biased to be "low entropy"/compressible/"short" in MP3/AAC— would it tend to generate sounds that sound "like sounds" rather than like white noise?
Recongizers are often used only on inputs from some nontrivial distribution (e.g. image models on actual images and not white noise). You need a model for that distribution to use a discriminator as a generator.
Compression is ~equivalent to a model of the thing being compressed in a closer sense than that for recognizers. It's just often not that good of a model, because, among others, runtime performance of decompression is a consideration that at some point overrides optimizing for smallest (expected) compressed size.
