Adaptive AI Speech Therapy and Real-Time Correction Engine

Speech disorders affect over 1.5 billion people globally, yet access to timely, affordable, and expert speech therapy remains severely limited, particularly in developing nations. This paper presents the Adaptive AI Speech Therapy and Real-Time Correction Engine — a fully automated, web-accessible intelligent platform that bridges the gap between traditional clinical speech therapy and scalable, technology-mediated language rehabilitation. The system integrates four synergistic modules: (1) Automatic Speech Recognition (ASR) via the Google Web Speech API achieving a Word Error Rate (WER) of 4.2% under clean conditions; (2) Word-level Pronunciation Comparison using Python's difflib SequenceMatcher, classifying each spoken word into one of four diagnostic categories — correct, missing, extra, or mispronounced; (3) Grammar Error Detection via LanguageTool NLP, attaining 92.6% precision and 85.1% recall across five grammatical error categories; and (4) a Weighted Performance Scoring algorithm combining pronunciation accuracy (70%) and grammar correctness (30%) into a clinically interpretable 0–100 composite score with five named performance bands. Implemented on a high-throughput asynchronous FastAPI backend with Pydub-driven format-agnostic audio preprocessing (supporting MP3, WAV, M4A, OGG, FLAC, WebM), the system delivers end-to-end feedback within 1.24–1.72 seconds. Experimental evaluation across 50 diverse audio samples spanning five speaker profiles, five audio quality conditions, and five sentence complexity levels demonstrates pronunciation accuracy of 95.8% on clean audio, F1-scores ranging from 0.764 to 0.954 across speaker profiles, and grammar detection performance surpassing all benchmarked commercial alternatives. The system's open-source architecture, self-hostable deployment model, and documented REST API collectively position it as a clinically viable, equitable, and extensible supplement to traditional speech therapy.