The art of the pause

Will a computer ever be able to synthesise speech that cannot be distinguished from a (good) human reader? Yes, I am sure it will - and soon - get very, very close, but I doubt it will ever quite get fully there. There is a fundamental problem. TTS is matching units of spoken sound to graphic symbols, and it then amends its selections to emulate the cadences of natural human speech according to the position in the sentence, adjacent words, and punctuation; it is getting steadily better at that as computing power and speed increase. However, it is making no judgements reflecting semantic content, only calculating probabilities from rules it has derived from samples. It gives the illusion of semantic understanding, but cannot independently actually encode emotions.

I once asked the consummate audiobook reader, the late Andrew Sachs, how he approached reading. "I want the writer to be noticed, not me. It's like the background music in a film. You shouldn't notice it, but it should have an effect on you. I use the pauses to indicate reaction. Audio is a sound medium, not a speech medium."