Straight out of a sci-fi book, Google Duplex is the application of natural speech and deep learning to create the most uncanny conversation imaginable.
Google Duplex is the brainchild of Google Voice Search and WaveNet, "a deep generative model of raw audio waveforms." The announcement underscored all of Google's efforts to make AI more functional, but nothing has quite struck a chord or shown a true glimpse of the future the way Google Duplex has.
The idea is to tell Google a specific action and an artificial human interface will be able to conduct the conversation with an actual human. That requires a lot of technology, especially with the syntaxes and nuances of the human language. For example, saying "OK for 4" might confirm either the number of people or the time of day. Or perhaps interrupting the A.I. to confirm a phone number.
That's why natural language processing is so hard to nail, but Google is certainly one step closer. On the technical end, and trust me this is very layman, Duplex is a recurrent neural network (RNN) designed to cope with these challenges, built using TensorFlow Extended (TFX). Basically, it can use the history of the conversation to make decisions.