it works, in an American context, because the states would have to take it to court and determine if this actually fell under the rules of making a recording.

Argument for, there is no recording stored anywhere thus it does not need consent.

Argument against, there is obviously a recording made. A recording is made and transmitted between devices and Apple and afterwards deleted, that this happens almost real time as the conversation does not change the fact a recording exists for some short time.

Probably Argument for wins, but assume at some point the transcriptions become nearly perfect then expect it to be litigated again.

It is "not stored" "not recorded".