It's different than just picking highly directional frequencies and using a tight beam. What they're doing is sending out two ultrasonic waves, one at a fixed frequency, and another variable frequency one near the first one's frequency. When the waves hit a surface (your head, a wall, or a table) the two waves undergo constructive/destructive interference. What you end up hearing is the beat frequency (the destructive interference one). The constructive interference wave will be well out of your hearing range, like the original two waves were.
Very neat. This can't be done well using standard methods, i.e. high power speakers with waveguides, as the (relatively) long wavelengths we can hear cannot be beamed well.
Likely, it sounds nothing like speakers or headphones. Not surprising since it works nothing like speakers or headphones. It probably won't be anywhere near hifi for a while either. If mapping out HRTF for headphones tough, this would be extremely difficult, since with headphones, you're dealing with a driver unit a fixed distance from the ear. With this, the equivalent driver unit would be the layer of air right next to your ear. Still, for low-fi uses, this thing is like magic.