NSF Convergence Accelerator Track H: Convergent, Human-Centered Design for Making Voice-Activated AI Accessible and Fair to People Who Stutter

Michigan State University, East Lansing MI

Investigators

Abstract

Disfluencies in speech are common artifacts in conversation, but they are especially prevalent in individuals who stutter, a community of more than 70 million people worldwide. It is well-documented that people who stutter consistently experience employment discrimination, diminished labor market outcomes, and societal stigma. The increasingly pervasive use of exclusionary voice-activated artificial intelligence (AI), which are designed, trained, and tested without considering communication that varies from societal norms, can act as a barrier to daily life participation and employment access for communities such as individuals who stutter. Worse, such technology can actively discriminate against people with speech differences in employment contexts. Therefore, there is an immediate and compelling need for efforts to reduce these barriers and empower people with communication differences and disorders to fully and equitably access all forms of speech recognition systems, including personal voice assistants, automated phone interfaces, and job-preparation and hiring software.This Convergence Accelerator project proposes a multidisciplinary, use-inspired approach that leverages cutting-edge advances in AI, as well as deep understanding of the nature and experience of stuttering, and the legal, ethical, and labor market implications of increased use of voice-activated systems. In partnership with many stakeholders, the project will develop and distribute high-impact solutions to a major national and global challenge: accessibility and fairness of voice-activated AI for disfluent speech. Improving the ability of voice-activated AI to appropriately parse and decode disfluent speech will increase quality of life, equality of opportunity, and access, not just for people who stutter, but also for other vulnerable populations and for society at large because all speakers are disfluent to some extent. The goal of this Convergence Accelerator project is to resolve limitations in voice technology by developing and implementing policy-, advocacy-, and AI-based solutions to make voice technology accessible and fair to people who stutter. The project will contribute to advancing knowledge through development of inclusive training and test datasets as well as annotation for accessible automatic speech recognition (ASR) and development of novel ASR deep learning models. Proposed research studies will establish a convergent and holistic understanding of how the nature and experience of stuttering impacts and intersects with AI accessibility and fairness in voice-activated technology, identify barriers and facilitators of access to existing voice-activated AI among people who stutter, and evaluate the effectiveness of guidelines and audit tools. Finally, this activity will engage a multidisciplinary and multisectoral network of partners to ensure participatory research design with a focused plan to recruit a wide range of participants, widespread dissemination of findings, and uptake of new, accessible technology. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.