The imec-IDLab-UGent speech team wins several tracks in two international speaker verification challenges

(04-12-2020) The imec-IDLab-UGent speech team won two international speaker verification challenges organized in conjunction with the Interspeech 2020 top conference.

In both the Short-duration Speaker Verification Challenge 2020 (SdSVC-20) and the VoxCeleb Speaker Recognition Challenge 2020 (VoxSRC-20) the team achieved first place, while competing with more than 50 academic and industry research teams from all over the world.

With these results IDLab demonstrates its expertise in text-independent speaker verification.

Congratulations to Jenthe Thienpondt and Brecht Desplanques, supervised by Professor Kris Demuynck.

Automatic speaker verification

Automatic speaker verification checks if the speech in two recordings was uttered by the same person. Deep learning has revolutionized this field by storm in 2017. Computers now outperform human experts by a large margin, making up to 15x less errors in certain conditions. Current systems distinguish voices with high accuracy from speech fragments as short as two seconds recorded in multilingual contexts such as interviews and movies.

Research at IDLab on self-supervised speaker verification allows large cost savings factor in the development of these systems by eliminating the need for labeled training data.

Speaker verification can for example be used to find audio fragments uttered by a public person in huge media archives, with current systems being capable of sifting through 2500 hours of audio in 1 hour. The technology is also used by personalized digital assistants such as Google Assistant and Amazon Alexa to quickly and reliably verify the user, securing access to certain services and making responses personalized. Other applications include call center voice authentication and the automatic detection of identity theft in voice phishing.