I'm wondering too. I had a glimpse at a paper and it seems it works because even though ultrasound is "inaudible to humans, still leaks into the recorded speech due to non-linear amplification of the microphone’s circuit". The chart thought shows the leakage has small dB (-20) in audible spectrum.
Too bad researches didn't leave any samples of recording in the repo. Curious how it would be effective after: 1) filtering frequencies to 0-16kHz 2) getting 1-2sec sample duration of noise profile while nobody speaks and using noise reduction filter from audacity