Looking into better compression algorithms inspired by ML, LLM techniques under different constraints. For example if the receiver has a prior knowledge because future devices will have some foundational models that already understand higher level concepts like how a human looks or speaks, do we really need to send the entire recording or just the "words", "intonation" etc and reconstruct it on device.
Fabrice Bellard and other has done some work in this domain https://bellard.org/