The best approach would be (only half-joking here) to wait a year and then use Baidu or a dozen other (Lyrebird included) group's voice synthesis neural nets (g

The best approach would be (only half-joking here) to wait a year and then use Baidu or a dozen other (Lyrebird included) group's voice synthesis neural nets (generally either a RNN or a WaveNet-style CNN). The last WaveNet samples were indistinguishable from real for me.