The model behind probably doesn't really check for individual notes and their roots. It "listens" like a human listens (people recognize songs when you hum a piece of song, even though it's not correct root/interval/perfect pitch) so your hum doesn't need to be perfectly music-theory-accurate.