An AI That Listens to Children Read is Transforming Literacy in India

India has one of the largest school systems in the world, but millions of children finish primary school unable to read a simple grade level text. According to the Annual Status of Education Report, only 42.8 percent of Grade 5 students in India can read a Grade 2 level text. A nonprofit called Wadhwani AI is attacking this crisis with a deceptively simple idea: an AI that listens to children read aloud and tells teachers exactly where each child is struggling. Its Oral Reading Fluency tool has already delivered more than 3.7 million assessments across 30,000 schools in Gujarat.

The tool, known in India as Vachan Samiksha, was built by Wadhwani AI, a nonprofit founded by Indian-American tech entrepreneurs Romesh and Sunil Wadhwani. It runs on automatic speech recognition models trained specifically on child speech in Indian languages, a critical distinction, because children read differently than adults speak, and most commercial speech systems are built for adult voices in English.

The Scale of the Reading Crisis

The Annual Status of Education Report, widely known as ASER, is India's most trusted household survey of learning outcomes. Its findings on reading have been stubborn for years: a large share of children in Grade 5, and even beyond, simply cannot read fluently. A child who cannot read a Grade 2 text by Grade 5 cannot learn from textbooks, cannot do word problems in mathematics, and falls progressively further behind with every passing year.

The traditional way to diagnose this is a one on one oral reading assessment, where a teacher listens to each child read a short passage and records mistakes. Done properly, it takes five to ten minutes per child. In a classroom of forty or fifty students, and with teachers already overloaded, it simply does not happen at scale. The result is that struggling readers are invisible until they fail.

How the AI Listens

Wadhwani AI's solution replaces the clipboard and stopwatch with a phone or tablet. A child reads a passage aloud. The automatic speech recognition engine, trained on recordings of children reading in Indian languages, transcribes and evaluates the reading in real time. It measures correct words per minute, flags missed words and substitutions, and identifies patterns in the errors, whether a child struggles with specific letter sounds, skips words, or reads too slowly.

The teacher then receives a dashboard of actionable insights instead of a raw score. The system recommends targeted interventions for each child, grouping students by the specific nature of their difficulties so that limited instructional time goes where it matters most. The assessment, which once took a trained evaluator minutes per student, is now automated and can be repeated regularly to track progress over the school year.

Three Million Assessments and Counting

The deployment in Gujarat demonstrates what scale looks like. The tool has conducted more than 3.7 million fluency assessments for students in Classes 2 to 8 in Gujarati, supporting 120,000 teachers across 30,000 schools. Those numbers matter because assessment is the foundation of instruction. Before this tool, a state could measure reading levels once a year through survey sampling. Now individual teachers can see how every child in their classroom is progressing, multiple times a year.

The project has also advanced the broader goal of Indian language AI. Building reliable speech recognition for child speech in Gujarati required large annotated datasets and deep cultural and contextual understanding, capabilities that also feed into the development of Indian language large language models. The solution won recognition through the MIT Solve Tools Competition, which highlighted its real time automated evaluation of oral reading and its teacher facing intervention recommendations.

Why Speech Recognition for Children Is Hard

Most speech recognition systems are trained on adult voices speaking fluently. Children learning to read do the opposite: they speak haltingly, mispronounce words, pause in unusual places, and drop word endings. A system trained on adult speech would flag nearly everything a struggling reader says as an error, or miss errors entirely.

Wadhwani AI trained its models on child speech in Indian languages, which required collecting and annotating large volumes of recordings of real children reading. The result is an engine that understands the difference between a hesitant pause and a skipped word, and that works in languages that commercial AI companies historically underserved. Work is now underway to extend the tool beyond Gujarati to English, Hindi, and Marathi.

A Model for Foundational Literacy

The deeper promise of this work is that assessment stops being a bottleneck. When measuring a skill is cheap and fast, the skill can be taught deliberately. Teachers in Gujarat can now identify the specific children at risk of finishing primary school unable to read, and intervene years earlier than the system previously allowed.

Foundational literacy is the gateway to everything else in education, and India is not alone in facing this gap. If the combination of child trained speech recognition, automated assessment, and teacher facing insights can be replicated across languages and countries, the AI that listens to a child read in a Gujarat classroom could become one of the most consequential education tools of the decade.