VAE objectives are useful for pushing embeddings into a Gaussian distribution.

Here's some work on low-latency neural compression that you might find interesting: https://arxiv.org/abs/2107.03312