The SPROUT (Speech Production Repository for Optimizing Use for AI Technologies) project is designed to support the development of automatic speech recognition (ASR) systems and AI-driven tools for early childhood assessment of speech and language development. SPROUT focuses on collecting speech samples from four-year-old children — a strategic choice that targets the developmental midpoint of speech sound acquisition in American English. This age group provides an ideal balance: minimizing the frequent developmental errors seen in younger children while still challenging model performance more than fully developed speech in older children. To ensure equity and robustness in AI development, SPROUT prioritizes speaker variability. The dataset is built using a balanced recruitment strategy across racial/ethnic groups (Black, Latine, and White) and socioeconomic status (at or below 100%, and between 100%–200% of the federal poverty line). Recordings were collected from over 300 unique participants across seven geographically diverse U.S. cities — St. Louis (MO), Los Angeles (CA), Dallas (TX), Orlando (FL), Atlanta (GA), Baltimore (MD), and Iselin (NJ)—through a partnership with a healthcare research firm. Importantly, SPROUT was designed with robust ethical protections to safeguard the use of child speech in AI research. All data collection followed rigorous human subjects research protocols, including informed parental consent, options for data sharing preferences, and strict data privacy and security measures. These protections ensure that SPROUT supports innovation in a way that is transparent, respectful, and centered on the rights and dignity of children and families. Developed through a collaborative, community-engaged approach, SPROUT enables cross-disciplinary research in speech-language pathology, machine learning, education, and child development. It supports the development of tools for early speech disorder detection, language development tracking, and culturally responsive assessment practices.

Acknowledgments: We extend our sincere thanks to the families who contributed to this data collection. We are also grateful to the undergraduate research assistants—Megan Delande, Hannah Ma, Lucy Madsen, Emily Park, Kevin Rha, Brookelyn Slonaker, and Sebastian Hureta—for their dedicated support of this project. Additional support was provided by the Roxelyn and Richard Pepper Department of Communication Sciences and Disorders in the School of Communication at Northwestern University, to which we are grateful.

Approved by the Northwestern Institutional Review Board: STU00219398