The concept seems quite interesting. I can hear slight differences between the different versions, but I honestly can't hear those differences as conveying any particular emotion. If you mixed them up and asked me which one sounded most "happy" or most "scared" or most "sad", I'd be stumped.
Nor can I very well. After some googling, I see that apparently this would be called "auditory affective agnosia" (if it were at the level of neuropsychological pathology). I wonder if this software could be used diagnostically in disorders that are associated with deficits in empathy or emotional identification -- autism, Antisocial or Narcissistic Personality Disorders, etc.
Perhaps it could also be used as a therapeutic tool. One could train with it in order to reduce the affective threshold at which one detects emotion.
> One could train with it in order to reduce the affective threshold at which one detects emotion.
Interesting idea. If software like this can smoothly adjust intensity, you could find the threshold at which you can reliably detect it, and attempt to push that downward to pick up more subtlety.
In the first video (at around 0:33) you can see the voice as if it was applied the "scared" filter too much, and cutting only the section where she says "yes, we met with X" out of context to make it sound like a shameful confession instead of the firm statement of known information that it is (0:41 of second video).
Amazing. Imagine combining this with facial emotion detection algorithms like those in Microsoft Kinect and making an AI that can automatically empathize with users. When you're happy, the AI is happy. When you hate the world it hates the world. Maybe you could even optimize the UI so that certain offers like buying things or rating apps would only be shown when the user is in a good mood.