UBA develops AI that listens to children read and measures reading fluency

At the University of Buenos Aires, in the research centers of the Faculty of Exact Sciences, a team of computer scientists and neuroscientists is building something quieter and perhaps more consequential than the generative AI systems that dominate headlines. They are teaching a machine to listen to children read.

The project began two years ago in the Laboratory of Applied Artificial Intelligence, where Juan Kamienkowski and his team—Jazmín Vidal, Octavio Castro, Luciana Ferrer, and Pablo Riera—took on a problem that sits at the heart of Argentina's educational crisis: most students leave primary school unable to read with fluency or comprehension. The researchers started with audio recordings collected by the province of Mendoza, where teachers had begun systematically capturing how children read aloud in classrooms. The goal was to train an AI system that could listen to those recordings and measure, automatically and instantly, how well each child was reading.

The challenge is deceptively simple to state and enormously difficult to solve. Reading fluency—the speed and accuracy with which a child can decode text—is a reliable predictor of reading comprehension. Research shows that no child should finish first grade reading fewer than 40 words per minute. Yet measuring this across thousands of students, identifying which ones are struggling, and intervening before they fall further behind requires both precision and scale. The AI system Kamienkowski's team is developing does exactly that: it listens to a child read, counts how many words they pronounce correctly, and flags the errors that matter.

Kamienkowski explained the work using an analogy to YouTube's automatic captioning. YouTube can subtitle videos reasonably well, he noted, but it makes mistakes. When the audio itself contains an obvious error, YouTube corrects it because the platform's goal is to normalize speech so everyone can understand it. The UBA team's task is different. They are not trying to correct what children say. They are trying to identify when children make mistakes—to distinguish between different types of reading errors and measure them precisely. The system currently calculates words read correctly per minute, the standard metric for fluency. Eventually, once trained on a broader range of error types, it will provide more granular feedback about what specifically is failing in each child's reading.

The obstacle is data. Machine learning algorithms require enormous amounts of training material, and obtaining recordings of children's speech is difficult. Obtaining those recordings with precise annotations—marking exactly where errors occur and what kind they are—is far more difficult still. The team has been working methodically through this problem, updating their models as they gather more examples from Mendoza's schools.

The next phase is to move from the laboratory to the classroom. Kamienkowski and his colleagues are developing an application that teachers can use directly: record a student reading, process the audio instantly, and receive a diagnosis of that student's reading difficulties along with suggested strategies for intervention. That application will eventually be available to any province in Argentina as part of the National Literacy Plan, allowing educators across the country to identify struggling readers early and act before gaps widen into failure.

The UBA team is not working in isolation. They are partnering with Dale!, an organization focused on reading comprehension, and the Fundación Natura, which helps source and annotate the audio data and coordinates with schools. Kamienkowski emphasized that the university's role is to develop the technology and train provincial staff to manage it independently—a transfer of capacity designed to make the system sustainable beyond the research phase.

In the same laboratory, the team is also developing a second AI system, this one designed to predict which students are at risk of dropping out of school entirely. Using data from Mendoza's education system—attendance records, grades, and other indicators of educational trajectory that teachers enter in real time—the AI generates a color-coded alert system: green for low risk, yellow for moderate concern, red for imminent danger. The system flags students before they leave, giving policymakers a chance to intervene. Like the reading fluency tool, this system is being designed for provinces to operate themselves, with the university providing the technical training to make that possible.