but can't a decision model or some other model deduce "P" or "B" based on the context (i.e., previous words said)?

reminds me of tesla fsd v. waymo. may be a market for both? i.e., camera-only or camera + specific hardware.

Not well enough! You can get to 15-20%ish WER but going beyond that doesn't seem feasible with the info available from a camera, even with a great language model