AI-POWERED SECOND LANGUAGE ACQUISITION FOR MINORITY LANGUAGES USING FLASHCARDS

Abstract

This project aims at filling the critical gap of interactive and adaptive resources for learning minority languages in general and Kazakh in particular. We developed a webbased language learning system that leverages state-of-the-art Speech-to-Text (STT) and Text-to-Speech (TTS) technologies to help English and Russian speakers acquire Kazakh pronunciation and vocabulary. The platform presents users with flashcards containing images of English/Russian terms and prompts them to pronounce the corresponding Kazakh word. A specialized STT module—using the pre-trained wav2vec2-large-xlsr-kazakh model—evaluates pronunciation accuracy, and a transformer-based TTS engine, trained on the ISSAI KazakhTTS2 and S¨oyle corpora, provides high-quality audio examples. A key feature of our platform is its adaptive learning algorithm, inspired by the SM2 spaced repetition system. Each flashcard maintains an “easiness factor” and review interval, which are dynamically adjusted based on the learner’s self-graded performance to optimize retention. We structured the solution into modular components—STT inference, pronunciation scoring (via Levenshtein distance), SM-2 flashcard scheduling, and TTS feedback—and validated each with unit tests covering critical functions. End-to-end System Integration Tests achieved a 95 % pass rate, ensuring seamless interaction among modules. In a User Acceptance Testing phase with 15 participants, our system received an average satisfaction rating of 4.5/5 and demonstrated a 30 % improvement in vocabulary recall after one week. Careful implementation (backend APIs, frontend UI), and comprehensive evaluation (unit/SIT/UAT), this project confirms the feasibility of applying advanced deep learning techniques to minority language learning and establishes a scalable framework for future extensions.