AUTONOMOUS CHATBOT FOR SECOND LANGUAGE ACQUISITION: AN AGILE, EDGE-BASED SYSTEM FOR EARLY CHILDHOOD EDUCATION

Abstract

This research addresses critical limitations in traditional second language acquisition (SLA) methods for young children, particularly the deficiencies in personalization and the dependency of modern AI solutions on persistent internet connectivity. Traditional SLA approaches often fail to maintain children’s engagement and provide real-time adaptive feedback, while existing AI-driven language learning platforms typically require cloud infrastructure, raising privacy concerns and excluding learners in resource-constrained environments. To address these challenges, we have developed an offline, responsive, and privacy-preserving autonomous chatbot system that operates on edge device. The primary objectives of this research include developing an offline interactive chatbot platform, enabling natural human-like language interactions tailored for young learners, implementing a personalization to fit individual learning progress, optimizing performance for the computational constraints of the Raspberry Pi 5, and ensuring comprehensive data privacy through local processing. Our methodology involved a pipeline architecture integrating carefully selected and optimized AI components, including Whisper for STT, a quantized LLaMA-3.2 (1B) for the LLM, and Piper for TTS. We also developed a web application for personalizing learning experiences. The main results demonstrate the feasibility of running a complete conversational AI system on a Raspberry Pi 5 with a Word Error Rate of 8.2% in quiet environments and an end-to-end response latency of 2.77 seconds. While hardware limitations prevented the implementation of image generation and aggressive model quantization without significant quality loss, the system showcases the viability of edge-based AI for accessible and privacy-centric educational applications, offering unique advantages over cloud-dependent solutions. Future work will focus on reducing response latency through further model optimization, exploring cost-effective hardware alternatives, expanding the educational scope with more advanced language models, and investigating techniques for enhanced multimodal learning experiences.