I do wonder if you actually need two models here. Audio-to-audio hindbrain on the client, and a beefy text-mode frontal lobe somewhere in the cloud, with the comms between them explicitly trained in as a potentially low-bandwidth steering connection transferring embeddings, not text.