I'm not familiar with how it works on the implementation side. A ML conference that I attended had a presenter working on this area and beating AI impersonation was their #1 priority (along with other trivial approaches).

For what it's worth, this might also be just a security theater from the banks, though.

The Natwest app will sometimes ask you to say words / numbers when doing a verification for e.g. a large transfer. Probably beatable by decent OCR feeding into voice synthesis with facial animation though (they don't have voice prints to check against that I'm aware of but even if they did...)