ID R&D takes top spot in international voice biometrics accuracy challenge
The latest results of the VoxCeleb 2023 Speaker Recognition Challenge are in, with ID R&D placing first in both the detection cost function (DCF) and equal error rate (EER) metrics.
VoxSRC 2023 Track 2 tests the effectiveness of speech-independent voice biometric authentication (i.e. Without passphrases). ID R&D submitted its IDVoice software for evaluation to achieve the top marks.
The challenge is organized by researchers at Oxford University, the Korea Advanced Institute of Science and Technology (KAIST), Carnegie Mellon University and AWS. It is designed to test the effectiveness of current voice biometric technologies on speech samples captured “in the wild,” according to the announcement from ID R&D. The dataset is composed of online videos of celebrity interviews, with a wide range of quality, speaker numbers and background noise conditions.
Track 2 assesses algorithms trained on an open dataset, rather than one defined by the competition organizers.
“VoxCeleb is a rigorous challenge requiring speaker recognition without passphrases during conversations in unpredictable environments,” comments Konstantin Simonchik, chief scientific officer and co-founder of ID R&D. “Our success can be attributed in part to bringing all available resources to bear to achieve the best results. It’s a particularly important win for ID R&D because it demonstrates the VoxCelebperformance of our technology under conditions that could be expected in real-world conversations, such as with emerging generative AI-powered chatbot technology.”
Other developers submitting algorithms include Unisound and Intema.
The results were announced at the VoxSRC Workshop at Interspeech 2023 in Dublin last week.
ID R&D was also recently granted a patent for a system to detect deepfakes and spoof attacks against voice biometrics systems.