Google Duplex is a tech developed by Google which uses AI to have natural conversation for specific tasks with people over phone.
In all of our Sci-Fi movies, shows and novels, a natural conversation between a human and a computer is a premise which is taken for granted. Be it Hal from 2001: A Space Odyssey to child genius Dexter talking to his secret lab’s computer, natural conversation is the ideal interface to interact with computers that we have always wished for.
Despite having tremendous advancements in the fields of natural language processing and deep learning, voice from computer has always seemed very rigid, cold and robotic and our conversation nothing beyond the hard coded responses.
To solve this problem Google gave demo of it’s latest technology called Google Duplex which is an AI tech which can talk to real people over phone asynchronously and in a fully automated manner to accomplish specific tasks.
It is nothing short of a true breakthrough in AI and human centric computer design wherein the machine adapts around the human and not the other way round. Currently it is capable of doing specific tasks such as booking an appointment with hair saloon, making restaurant reservations, getting information about business not mentioned on their website like opening hours etc.
Although the system is highly advanced and sounds extremely natural while doing specific tasks in a closed domain, but it is still not capable of doing free flowing general natural conversation.
Duplex is uses a Recurrent Neural Network built on top of google’s open source machine learning library TensorFlow extended. It was trained on massive corpus of data from anonymous calls and conversations.
To make it sound natural it uses a concatenative Text To Speech model and a synthesis TTS to do circumstantial intonations in the speech. Interestingly you can hear actual speech disfluencies such as ‘uh’s and ‘umm’s as the system does processing and at the same time these disfluencies make the voice sound even more natural. And to manage latencies (100ms or less) to uses multiple approximation algorithms, in some cases intentionally increasing latency to sound more human like.
Google Duplex is fully autonomous and is highly capable of completing specific tasks it is designed for, but in case if the situation is unusually complex, then it informs the human user to take it from there.
There are vast majority of businesses who don’t have any kind of booking system installed and do all of their business over phone, google duplex can really add great amount of value for them.
Its great to see how far we have come on building a true conversational AI and Duplex is surely a giant leap in that direction.