In the initial analysis phase Voice examines English texts and calculates the overall probability of each phonetic being spoken. It also analyses - for each phonetic - which of the other phonetics are more likely to precede and follow it.
In the articulation phase - during exhibitions - it applies the overall probability table to bias a random generator and constructs a seed sentence. It speaks this sentence through a bank of 44 speakers. Each speaker being associated with one particular phonetic.
Using the proximity probability tables, it then rearranges the phonetics in that sentence into more comfortable positions. It speaks the sentence again. It does this until it gets bored and then generates a completely new seed sentence.
My thanks to Ania Babiek for her voice.