Beyond User-centric: Modelling Privacy and Fairness Effects of Speech Interfaces on Community- and Society-Levels

Research on privacy protections for speech technology has focused on user-centric approaches. While this is a well-motivated starting point, broader AI research includes community- and society-level effects, known as human-centric AI (HCAI). This work addresses human-values-centric speech technology (HVST) through two use cases: 1) speech-operated devices controlled by one but used by several users, and 2) telecommunication services for speech without interoperability with other services. Our experiments model user preferences and choices regarding privacy and utility. The community-level analysis evaluates how users balance their preferences with those of their peers. The society-level analysis assesses the proportion of users oppressed and discriminated against by their peers. The main results indicate that oppression and discrimination due to speech technology can be reduced by allowing fine-grained tuning of privacy preferences and mandating interoperability between communication platforms.