2026-07-17 –, Johnson Great Room
Improving automatic speech recognition (ASR) for children is needed to enhance education and early childhood development. When ASR fails for children, reading assessments mis-score, speech therapy tools become unreliable, and many classroom tools cannot be built at all. The ASR gap exists because data sensitivity complicates the collection and sharing of transcribed child audio.
In this presentation, we’ll share how to unblock progress by creating public useful AI infrastructure even when data can’t be shared openly. We’ll discuss what makes child ASR so hard, how we advanced the field with an AI modeling competition, and best practices for sharing pretrained models.
Better child-centered automated speech recognition (ASR) is needed to unlock new research and tools to help teachers teach and students learn. ASR is largely solved for adults but remains a challenge for children, especially in noisy, real-world learning environments. Children’s speech presents distinct modeling challenges: greater acoustic variability, inconsistent pronunciation, uneven speech and linguistic development, and unpredictable grammar and vocabulary. There are also wide differences across age, accents, and speech tasks. Yet the development of robust ASR models for children is fundamental to universal screening, personalized literacy and reading instruction, speech therapy, and educational games. The applicability of these models extends to communication, commercial, and medical contexts.
This talk will present new, benchmarked ASR models developed through a crowdsourced AI competition. The competition draws on a combined corpus of pre-existing child speech datasets and newly curated, annotated recordings, comprising 560,000 transcribed utterances and 519 hours of child speech. We will discuss how hosting a competition can enable progress in a domain where the data is sensitive, difficult to collect, and difficult to share.
Then, we will break down what we learned from the competition about effective modeling approaches for child ASR. We will discuss the strengths of various transformer-based architectures, such as Parakeet, Canary, Whisper, and Qwen, as well as fine-tuning strategies to produce transcription outputs suitable for diagnostic and speech-screening applications. We will also discuss why competitions remain useful in the age of AI agents.
Beyond modeling results, we will describe the broader AI infrastructure challenge at the center of this work. Improving child ASR requires access to large, representative datasets, but children’s speech raises difficult questions around privacy, identifiability, and responsible model release. We will discuss the tradeoffs involved in using private data to evaluate public approaches, collecting demographic information to assess bias while limiting privacy risk, and deciding what parts of an AI system can be made shareable when the underlying data cannot be fully open.
The goal of this talk is to present a pathway to unblocking progress by creating pre-trained models as a public good when the underlying data cannot be shared. The presentation is suitable for anyone interested in ASR and its use in educational contexts, as well as people in any field working with sensitive data, thorny data ethics questions, or the challenge of building shared AI infrastructure when datasets cannot simply be released publicly.
Katie Wetstone is a data scientist with a passion for leveraging machine learning tools to promote sustainable, ethical, and just change. At DrivenData, she works to implement open-source machine learning competitions and direct consulting projects that support mission-driven organizations. Her projects have spanned a variety of issues including public health, conservation, and education. She holds a BA in chemistry from Harvard University, and a Masters of Development Practice from the University of California, Berkeley.