Unrealistic Feedback in Socially Prescriptive Speech Technologies

Abstract

Socially prescriptive speech technologies (SPSTs) are technologies that claim to provide feedback to speakers about how other humans perceive their speech and communication style, based on algorithms trained on a set of static, prescriptive standards. Such systems are increasingly integrated into popular videoconferencing programs like Zoom and Microsoft Teams and are being used to score and judge employee performance across conversational contexts and domains. The current study evaluates the design, promises, and performance of two such systems, the Zoom revenue accelerator (ZRA) and Read AI, and discusses the theoretical and practical limitations of software that evaluates speech based on opaque standards of idealized speech. The paper then presents the results of a laboratory experiment designed to test whether the systems provide fair and realistic feedback to a diverse group of American English speakers. Results of linear mixed effects regression (LMER) models testing scores for “sentiment” and “engagement” indicate that both programs systematically provide lower for non-White speakers. Additionally, for sentiment, Read AI shows a bias toward L1 English speakers, while the ZRA shows a bias against L1 speakers. The systems also claim to evaluate the use of fillers, speech rates, and offensive speech, but they do by employing opaque metrics and coarse standards not based on psycholinguistic or sociolinguistic research. The ZRA rated 99% of speakers in this sample as employing too many fillers, and 78.4% as using a speech rate that was too fast. Read AI negatively evaluated 58% of participants who self-identified as neurodivergent, on the basis that the speakers used terms that the system flagged as “non-inclusive” language. Overall, the results of this study show that SPST systems do not provide realistic, contextually appropriate feedback to users. This study acts as a first step in understanding the mechanisms and standards employed in SPSTs, and provides a discussion of the linguistic, social, and ethical limitations of such systems as their use expands across employment and educational contexts.