Local models just aren't good enough for this. Not in a device that has to be energy-efficient, fit in a pocket and conserve its storage for other things the user may want to do with it.

Unless you have a homelab (which is just cloud under a different name), there's no way of doing a reasonable job here without sending your data to a third party.

In this case, preserving privacy and anti-discrimination are goals that you just can't reconcile.

What use case can't be accomplished by a local model. They can read things out loud and identify what you're about to walk into.