This New Tech Puts AI In Touch with Its Emotions—and Yours

A brand new “empathic voice interface” launched at the moment by Hume AI, a New York–based mostly startup, makes it potential so as to add a spread of emotionally expressive voices, plus an emotionally attuned ear, to massive language fashions from Anthropic, Google, Meta, Mistral, and OpenAI—portending an period when AI helpers could extra routinely get all gushy on us.

“We concentrate on constructing empathic personalities that talk in methods folks would communicate, quite than stereotypes of AI assistants,” says Hume AI cofounder Alan Cowen, a psychologist who has coauthored various research papers on AI and emotion, and who beforehand labored on emotional applied sciences at Google and Fb.

WIRED examined Hume’s newest voice expertise, referred to as EVI 2 and located its output to be much like that developed by OpenAI for ChatGPT. (When OpenAI gave ChatGPT a flirtatious voice in Might, firm CEO Sam Altman touted the interface as feeling “like AI from the movies.” Later, an actual film star, Scarlett Johansson, claimed OpenAI had ripped off her voice.)

Like ChatGPT, Hume is much extra emotionally expressive than most typical voice interfaces. For those who inform it that your pet has died, for instance, it can undertake an appropriate somber and sympathetic tone. (Additionally, as with ChatGPT, you’ll be able to interrupt Hume mid-flow, and it’ll pause and adapt with a brand new response.)

OpenAI has not mentioned how a lot its voice interface tries to measure the feelings of customers, however Hume’s is expressly designed to try this. Throughout interactions, Hume’s developer interface will present values indicating a measure of issues like “willpower,” “nervousness,” and “happiness” within the customers’ voice. For those who discuss to Hume with a tragic tone it can additionally decide up on that, one thing that ChatGPT doesn’t appear to do.

Hume additionally makes it simple to deploy a voice with particular feelings by including a immediate in its UI. Right here it’s once I requested it to be “horny and flirtatious”:

And when advised to be “unhappy and morose”:

And right here’s the significantly nasty message when requested to be “offended and impolite”:

The expertise didn’t all the time appear as polished and smooth as OpenAI’s, and it often behaved in odd methods. For instance, at one level the voice all of a sudden sped up and spewed gibberish. But when the voice could be refined and made extra dependable, it has the potential to assist make humanlike voice interfaces extra frequent and different.

The thought of recognizing, measuring, and simulating human emotion in technological programs goes again many years and is studied in a discipline referred to as “affective computing,” a time period launched by Rosalind Picard, a professor on the MIT Media Lab, within the Nineties.

Albert Salah, a professor at Utrecht College within the Netherlands who research affective computing, is impressed with Hume AI’s expertise and lately demonstrated it to his college students. “What EVI appears to be doing is assigning emotional valence and arousal values [to the user], after which modulating the speech of the agent accordingly,” he says. “It’s a very fascinating twist on LLMs.”