Google’s annual developer conference kicked off today and one of the most impressive demos was a new product known as Google Duplex. Part of the Google Assistant features, Duplex leverages years of investment from Google in natural language processing, deep learning and text to speech.
The on stage demo of the technology by CEO Sundar Pichai was met with applause from the crowd, matched with positivity by international media since. As impressive as what we seen is, we need to pause for a second and analyse exactly what we seen.
Duplex shows a recording, note, not live demo, of a conversation between a human and a computer, without the human knowing they were talking with a computer. On the surface sounds like a successful pass of the turing test.
After asking Google Assistant to make a hair appointment, Duplex takes over, looks at your calendar, understands you want just a haircut, not a blow dry, not a colour or a rinse or a perm, presumably because you’ve told Google this somewhere, in an email perhaps, although that seems unlikely. The back and forward negotiation around appointment availability is not a process any of us enjoy, so it was actually a great subject for an example.
That’s the thing, the 2 examples were just that, they weren’t live demos and the reason for that is that these were cleary two of the best phone calls (from both computer and human sides) out of hundreds or even thousands of calls, many of which undoubtedly went badly.
Remember when were were all sold on Siri, imagining that it could answer every question we ever had about the world, then in reality it often returned a web search when it didn’t know. Imagine that, but in an actual conversation and by an AI Duplex reporting to be you.
I want this technology to work, its potentially transformative for the use of our time and would have plenty of applications both in our personal and professional lives, but don’t expect it on your device any time soon. The Google Duplex recordings were two of the most natural sounding computer-generation voices we’ve heard, which is encouraging.
With more and more services like hairdressers, doctors, restaurants all making appointment making available on the web, its seems like an awful lots of effort has been poured into the voice side of the equation when human conversations to book appointments is on the decline. What would be useful is if the AI could visit a website, complete a form, book an appointment and let you know when its done. The problem with that is it looks an awful lot like behaviour from a bot and would likely be blocked by most sites.
If you missed the Google I/O 2018 keynote, there’s a recap video from Google below, skip to around 3:10 for the Duplex stuff.
In this regards a simple program that says “hello” also pass the turing test.