Picking up on tone of voice, body language, what’s left unsaid; these are all part of being what’s called a good listener.

Not everybody has the wit or subtlety or emotional intelligence to take such non-verbal cues: we all have a friend or colleague for whom the penny rarely drops, who usually can’t read the room.

The tin-eared could be about to get shown up even more: research carried out the Max Planck Institute shows that artificial intelligence systems could be within range of the best of us when it comes to listening and sussing out feelings. 

“[M]achine learning can be used to recognize emotions from audio clips as short as 1.5 seconds,” said Hannes Diemerling of the Max Planck Institute for Human Development, who said the bots showed “an accuracy similar to humans” when it came to perceiving joy, anger, sadness, fear, disgust and neutrality.

Diemerling and colleagues came to their conclusions after devising tests to see if machine-learning models could “accurately recognize emotions regardless of language, cultural nuances, and semantic content.”

Using Canadian and German datasets, they came up with “nonsense” sentences in clips cut to a second and a half, which they say is “how long humans need to recognize emotion in speech” and the briefest viable time-span to control for “overlapping of emotions.”

The team conceded that their investigation, which was published in the journal Frontiers in Psychology, had “some limitations.” They said “actor-spoken sample sentences” possibly do not “convey the full spectrum of real, spontaneous emotion” and described challenges posed by using different languages and datasets.

Follow us on Twitter