I think its a real persons voice. When i was learning portuguese, someone posted a interview with the lady who's voice was used in dl pt. I recon its the sysnthesis algorithm that needs tweaking for specific cases like this.
Fair enough, but I still think it would be simpler to just record a person saying each sentence and word individually instead of trying to fake it. I realize that probably takes up way more storage space but I stand by my point.
The Esperanto course does this, and the quality difference is incredible. Even without such recordings though, as long as one remains vigilant in learning around the shortcomings of the TTS, one should hopefully be okay.