AI trained to single out a voice in raucous crowd

A raucous cocktail party can be a nightmare for virtual home assistants like Amazon’s Echo and Google Home for their inability to pick out a single voice from the crowd and make sense of any instructions.

Researchers at the Mitsubishi Electric Research Laboratory in Cambridge, Massachusetts, have developed an AI that employs a machine learning technique called “deep clustering” to separate individual voices in real time, an attribute that extends the realms of automatic speech recognition, notes New Scientist.

The system, unveiled at Combined Exhibition of Advanced Technologies show in Tokyo this month, delivers 90 percent accuracy and can prove vital in improving smart assistants in homes and cars.