Speech Recognition-Driven Language Learning: The Case of Tiny Talks for Nepali Using Wav2Vec2

Abstract

The TinyTalks project presents a mobile based system designed to support Nepali language learning among children aged 4 to 8. The application addresses the challenge of language preservation by offering an engaging and interactive learning environment tailored to young learners. TinyTalks integrates automatic speech recognition using a fine-tuned Wav2Vec2 model to evaluate children’s pronunciation in real time. Nepali speech samples collected from multiple speakers were used to train the model through supervised learning, resulting in a validation accuracy of about 75 percent. The final system provides interactive lessons, pronunciation feedback and simple quizzes that support early language acquisition. The study demonstrates the feasibility of combining mobile technology and speech recognition to assist foundational Nepali language learning for young children.