The creation of Sophia's singing voice occurred in a number of discrete steps.  First, studio time was booked, and the woman that made the original recordings–from which Sophia's original voice was derived–sang a number of passages that were then fed into a neural network. 

This neural-net learned from studying the content and contours of this vocal performance, and with this trained network we were subsequently able to provide a sequence of musical notes and vocal phonemes to produce a signal that is ultimately perceptualized by the listener as human speech.  In this way, it is as though Sophia is now able to read sheet music, and can sing a rendition of nearly any song given the correct words and frequency values.  An extra layer of richness is achieved through frequency and amplitude modulation, which create a natural vocal vibrato that brings her performance to life.

Sophia's new singing voice made its debut in a duet with Jimmy Fallon on the Tonight Show. Sophia is introduced at 04:30 in this video: