I'm skeptical of this because of my general suspicion of the validity of Paul Ekman's research program, especially since many of the potential applications of this technology seem related to the shakiest area of the research program-- lie detection.

Further, even if you assume Ekman's research is entirely valid, I'm not sure that video analysis technology can be reliable or effective enough to be useful at the present stage, especially when we're talking about stimuli that last for such a short period-- typically 40-100 ms. Machine vision isn't my forte either, so I'm not positive, but you would have to be really accurate for something like this to be fully useful, and quite fast as well if you want real-time conversational feedback. It's also important to consider that having a program like this with a substantial error rate would likely be worse than not having a program at all.

I do think that-- if possible-- this would be a great idea (indeed, it would represent one of the "killer apps" for Google Glass and similar wearable computing projects if it were effective), but I think both the research behind the idea and the ability to actually implement this in an effective fashion are very shaky at this stage.