Socially prescriptive speech technologies: Linguistic, technical, and ethical issues

Speech technology tools can be powerful and transformative for individuals, businesses, and governments. Socially prescriptive speech technology (SPST) systems are advertised as helping speakers improve their speech and communication, especially in business contexts. These systems, such as Read.AI and Zoom Revenue Accelerator, use automatic speech recognition and large language models to give prescriptive feedback about a user's speech style during videocalls. However, such systems face significant technical, social, and ethical challenges related to providing feedback on naturalistic speech. SPST systems are prone to similar issues of algorithmic bias as other types of generative and automatic speech recognition systems for three reasons: (1) bias in training data and model architecture, (2) limitations related to idealized speech models and (3) lack of sociolinguistic context and one-sided evaluation of speaker-listener dynamics. As a result, these technologies have the potential to distort listener perceptions of speech and to increase unfairness in speech evaluation by both machines and humans. This article leverages findings from phonetics and sociolinguistics to explore the current state of SPST speech evaluation. It also articulates their potential uses and harms to inform researchers and the public about the functions of these systems as their use becomes more widespread across social domains.