Apple's long-awaited podcast data-transparency reveals an ad-listening audience with no clear format-preferences

I assume she has a canonical format for time (and it is almost definitely the one that NSFormatter uses for time intervals). It most likely attempts to use the largest unit it has one or more of first. While it doesn’t actually sound like a natural human, it is probable the closest you can get to what humans would do with the least rules. So while she replies differently when you say “90 seconds”, I bet she would say the exact same thing for “one minute thirty seconds” (and if you says “one and a half minutes” she would still say “one minute thirty seconds”…)