Many of those systems are configured in a way that they connect you to a human if they can't figure out what you're saying. The audio from those calls is fed t

Many of those systems are configured in a way that they connect you to a human if they can't figure out what you're saying. The audio from those calls is fed to a team of people who are responsible for teaching the system, so that the next model can be more accurate in directing calls. I once met someone who had been working that gig while in university. Apparently the vast majority of the recordings are boring and easy enough for a human to classify, but some are just people rattling off swear words, screaming, etc.