Skip the Small Talk: Google Duplex and Conversational AI

If you’ve ever used a voice-prompt telephone menu, you understand how frustrating such menus can be. You clearly say “Yes” or give the simple command “Tech support,” only to be met with a robotic voice telling you “Sorry, your response was not recognized.”

What if your phone could speak like a real person to other, live people?

Enter Google Duplex, which allows you to have Google Assistant make a phone call and actually talk to another individual.

Hands down, the biggest development over the last year has been Google Assistant. Used on over 500 million devices, it’s believed Assistant will support over 30 languages in 80 countries by the end of 2018. And forget mechanical, robotic voices. Instead of recording every possible sound, the company is looking to artificial intelligence (AI) to build a complete voice model.

That’s where Google Duplex comes in.

Demonstrated at this year’s Google I/O, Google’s annual developer conference, Google Duplex is jaw-dropping. The 60-foot screen showed line-by-line playback of Assistant making a call to a hairdresser and booking an appointment. The person taking the appointment likely didn’t even realize she was speaking to a computer.

That’s because Assistant speaks naturally to the individual at the other end of the phone, pausing and using filler words such as um and uh when answering. Complete with the idiosyncrasies often found in human speech, Duplex uses AI to accomplish tasks over the phone.

How is Duplex significant?

Developers’ goal is to make conversations sound as natural as possible and for Assistant to think in real time. Duplex is groundbreaking because it overcomes three main challenges:

• Getting the response recognized is one thing; ensuring that voice recognition software can distinguish nuances of language, accents and endless turns of phrase is another. Any and all of these can turn natural language processing on its ear.
• Not all conversations go the exact same way. Simple tasks such as booking an appointment or ordering food require high accuracy over a large dataset.
• Conversational AI must be capable of understanding context. People can say myriad things, and it’s impossible to predict—or construct—a conversation two or three levels deep.

Duplex must answer and sound like a person, talking fast and less formally. Duplex must also be able to understand context in a particular situation: it must be able to “think.” And think it does!

Duplex offers a glimpse into just how powerful conversational AI can be when designed for a narrow purpose. It offers both fascinating and scary implications.

Weigh and let us know what you think about the potential of Duplex. Leave a comment for us on Facebook or Twitter.

Or you could just get Duplex to call us with an opinion.