Detecting and Mitigating Bias in Voice Activated Technologies

In this talk, Wiebke (Toussaint) Hutiri shares her work on FairEVA, an open-source project to gain insights and build tools for mitigating bias in voice biometrics. The project is supported by a Mozilla Technology Award for advancing trustworthy AI and motivated by research that Wiebke will present at ACM FAccT 2022 in Korea. Speaker verification is a voice-based biometric that identifies speakers from their voice. It is used in voice assistants, call centers, and forensics. There are many parallels between the history and technologies of face recognition and speaker verification. However, while bias and discrimination are recognised as significant challenges in face recognition, little attention has been paid to them in speaker verification. This is a problem, as speaker verification systems are increasingly used in sensitive domains (e.g., financial and e-wallet services such as Safaricom where ‘my voice is my password’, pensioner proof-of-life verification, health and elderly care). In addition to sharing insights on identifying bias in speaker verification, Wiebke will also discuss the project’s advances on design guidelines for inclusive speaker verification evaluation datasets, their open-source library for evaluating bias and a technology audit that they are designing.

Presentation slides

Speaker Details

Wiebke is a PhD candidate at Delft University of Technology in the Netherlands. She graduated with a MSc in Computer Science and BEng in Mechanical Engineering from the University of Cape Town. Wiebke is working on trustworthy edge intelligence, with a focus on bias in voice-activated systems. She is the recipient of a Mozilla Technology Award, which she received to advance her work on fair voice biometrics. Wiebke believes that building local technical talent is critical for Africa, and has spent over a decade mobilising for change in South Africa’s engineering sector through Engineers Without Borders, initiating transformational projects like the Data for Municipal Infrastructure Assets programme. Wiebke co-organised the ML at the Edge workshop at 2022’s Deep Learning Indaba in Tunis.