Trained and optimized for fluent speech, speech AI works poorly for people who stutter (PWS): it often cuts them off from speaking and interprets the speech of PWS with four times higher error rates than average. The increasing deployment of voice AI in automated phone menus, AI-conducted job interviews, and everyday devices poses tangible risks to PWS.

Even when automatic speech recognition (ASR) systems do manage to transcribe stuttered speech, the resultant transcriptions often remove disfluencies like filler words, inevitably stigmatizing stuttering and denying the option for PWS to have their disfluencies preserved and normalized in transcripts.

To address the fluency biases embedded within today’s speech AI technology, we turned to our community and engaged in grassroots AI efforts with and by the stuttering community. Together, we develop datasets, metrics, tools, and techniques to measure, understand, and reduce fluency biases in existing ASR models.

Our working process with the grassroots stuttering community also showcases an alternative model for AI development, a model that builds and amplifies capacities within marginalized communities to challenge the existing concentration of AI power, allowing us to envision a technological future that server broader public interests rather than the profits and dominance of a few.

Publications

Datasets, tools, and models

All the following technical assets were produced and maintained by people who stutter.

Grants

We thank the following funders for support our work on fair speech AI for people who stutter:

Press