HSS statistically analyzises Top 30 hits to determine mathematical similarities in pitch, tempo, and structural format, as well as spectral analysis and over 25 additional variables directly from the digitalized audio signal.
It has been noted that there are a finite number of clusters that all most all hit songs fall into.
Although I have not had direct access to this data, it can be hypothesising that there are 18-55 mathematical archtypes based on the clustered formation.
HSS does not take into consideration of vocal content although the tonal aspects make a marked expression as a vocal insturment. While Vocaloid may still be in it's infancy, I believe a synthetic vocal insturment will be more efficent at matching the model of the mathematical archtype.
This voiceset sampling technique is similiar to MBROLA synthesis. Voice set creation is extensive and requires a studio enviroment. the possiblity of reanimating a deceased performance artist vocals is highly unlikely with this technique. Leave it to Yamaha to dress it up fun enough for kids to play with.
Check out this topic bending HSS link:
http://www.hitsongscience.com/index.php?p=w