The question is, is the multisecond latency because of processing and code efficiency limitations for a academic research project, or is it because the data is unusable without two seconds of smoothing? Given what I see in the video I could argue either way. I do note the latent signal is still a bit noisy, but then, touch screen input isn't necessarily clean either.

But it's also worth keeping in mind this is all off-the-shelf hardware. It seems very likely to me that if a cell phone or smart watch was designed to do this from the get-go that several easy hardware improvements and maybe a bit of custom DSP work would make this work much better. (By "easy hardware improvements", I mean things like speakers intended to emit frequencies for sonar, microphone arrays intended to receive them, etc.) From that perspective, even if the system we saw is fundamentally limited I'd still call it incredibly promising considering the constraints it is operating under!

If I were a smart watch manufacturer I'd be falling over myself to get one of my best engineers and one of my best recruiters an appointment with these people.

The accompanying paper[1] claims the phone lasts four hours running the current version of fingerIO, but also that improvements could be made to preserve power (such as reducing the sampling rate).

[1] http://fingerio.cs.washington.edu/fingerio.pdf

They are calculating autocorrelation and a Fourier transform so they need to buffer the data. Two seconds is probably the shortest buffer that works reliably.