Leveraging AI-based reading assessment to bridge foundational literacy gaps

• In Event: Developing and adapting voice AI for inclusive early grade reading assessments: insights from Sub-Saharan Africa and South Asia

Proposal

Foundational literacy remains a global priority, yet millions of children continue to struggle with acquiring basic reading skills. Traditional assessment approaches, paper-based, labor-intensive, and infrequent often fail to provide teachers with timely or detailed insights into learners’ needs. This applied research project introduces and evaluates an AI-enabled reading assessment application designed to overcome these challenges by delivering real-time, fine-grained analysis of children’s reading performance.

The intervention integrates speech recognition, acoustic modeling, and natural language processing to analyze children’s oral reading in multiple local languages. The system is trained on 200,000 annotated child speech samples, enabling it to handle the irregularities of young learners’ voices as well as diverse accents and dialects. A layered error-classification algorithm identifies not just whether a word was read correctly, but how it was misread, capturing omissions, substitutions, insertions, hesitations, and mispronunciations. These outputs generate individualized mistake profiles that distinguish between decoding challenges, phonemic awareness gaps, and comprehension issues. Unlike conventional assessments that yield only a score, this approach provides actionable diagnostic insights for both teachers and learners.

The tool operates offline-first, with local processing ensuring functionality in low-connectivity settings. Teacher dashboards synthesize data at both the learner and classroom level, enabling immediate instructional responses while reducing assessment burden. Aggregated system-level data further highlight geographic and linguistic inequities in reading achievement.

A crucial foundation for this application was the large-scale collection of children’s reading data across multiple languages and contexts. This dataset was used to train the AI models for robustness in low-resource environments, ensuring that recognition accuracy and error classification remained equitable across diverse learner populations. Without this groundwork, the app’s fine-grained mistake profiling and adaptive feedback loops would not have been possible.
The research follows a design-based methodology, with iterative pilots in multilingual communities. Early results indicate that technology increases assessment frequency, enhances diagnostic reliability, and uncovers systemic learning barriers invisible to traditional tools. For example, error analyses have exposed widespread difficulties in letter-sound correspondence tied to local orthographies, insights now being used to inform both pedagogy and policy.

This project illustrates how applied research and AI-driven tools can advance equity by uncovering hidden learning gaps and providing teachers with actionable pathways for remediation. By grounding innovation in contextualized data, the intervention shows how technology can be a mediator of justice rather than exclusion. At the same time, the work raises essential considerations around child data privacy, linguistic inclusivity, and algorithmic fairness. These reflections underscore that while technology is not a panacea, when responsibly designed and deployed, it can illuminate learners’ needs, empower educators, and strengthen the foundations of more equitable education systems.

Authors

• Nishant Baghel

• Vani Shree, Pratham International