Linguists take it as axiomatic that speakers are experts on their languages, both in grammar and usage. However, as Large Language Models (LLM) trained on text and speech become ubiquitous in domains from daily tasks to education and employment, human expertise about language is increasingly devalued. This talk will present the results of three studies that focus on how LLMs judge and purport to “fix” the speech of human talkers, also known as Social Feedback Speech Technologies (SFSTs). The first study shows how the Amazon Halo, a wearable device that claims to evaluate “tone of voice” does not function as advertised, and in fact systematically negatively evaluates the speech of Black talkers.  Results of the second study, which focuses on Read.AI and the Zoom Revenue Accelerator in videoconferencing contexts, describe how SFSTs reinforce narrow “standard” language ideologies and fail to provide actionable, realistic feedback to users. These systems also provide systematically worse evaluations for black speakers, as well as those who are neurodivergent. Finally, the third study analyzes the outputs of “accent translation” programs marketed by companies such as Sanas and Krisp, showing that such programs do not functionally “translate” accents but rather transform speech to an imagined “American” variety that is phonetically unnatural. Taken together, the studies show that “AI”-based programs that purport to evaluate human speech do so without consideration of linguistic principles or acknowledgement of speakers’ sociolinguistic competencies. Such systems also act without transparency for both designers and users by design, reproducing social stereotypes inherent to their training data. As a result, they advise humans to produce unnatural speech, and they punish speakers who do not conform to the narrow targets established by an LLM’s training data. As such technologies are already being used to make employment decisions, provide speech therapy, and even draft police reports, the fact that these systems systematically misevaluate speech represents a significant threat to all human speakers, most especially those from marginalized groups. 

Location: Royce Hall 156