Automatic speech recognition tools don’t work for children. Their voices, speech patterns, and phonetics are all different from adult voices. 

This is a major barrier to improving learner engagement, enabling scalable assessments, and providing broad coverage for screenings. With good ASR tools, children would have access to LLM models that need text to operate. Assessments for oral reading fluency, concept understanding, and literacy could be conducted automatically with immediate feedback. Initial screenings for speech pathologies could help direct more students to the interventions they need earlier. 

We’re launching a challenge to build the best state-of-the-art algorithms for ASR for children’s speech. These models will capture both what children say and how they say it. The goal is to provide open access to the best approaches. This means open data, open code, and open-weight pretrained models for use by a broad community of educators, nonprofits, and ed-tech companies.

If you have insights for our challenge design, please use the survey link above. Otherwise, sign up to get updates on the challenge!