Kutti AI: Voice-first offline companion detects real-time struggle for visually-impaired kids

Most educational tech relies on visual interfaces, excluding an estimated 1.4 million blind children and millions more with low vision. Kutti AI, presented in a new arXiv paper, flips that model entirely: audio is the primary and sufficient interface. Children learn curriculum concepts through spoken conversation, respond by speaking, and receive spoken feedback—no visual elements required. The system runs on commodity mobile hardware and is designed for low-connectivity settings common in underserved communities, with an on-device automatic speech recognition (ASR) model enabling full offline capability.

Kutti AI introduces three practical mechanisms for adaptive learning. First, a multi-signal struggle-detection engine combines response-latency analysis, wrong-attempt tracking, and keyword-based hesitation detection to decide in real time when to offer hints or simplify a question. Second, a multi-layered cross-language answer-matching pipeline uses translation, transliteration, Levenshtein-based fuzzy matching, and text normalization so children aren't penalized for code-switching or pronunciation variation. Third, the offline-first speech pipeline ensures usability without internet. A hackathon prototype supporting English and Tamil demonstrates the concept. The paper outlines a path toward formal evaluation, but already illustrates how a small, carefully-engineered voice-first system can dramatically lower both accessibility and financial barriers to early education for visually-impaired children.