Suffix-Aware AI-Driven Pronunciation Evaluation for Kannada Language Learning in Diaspora Student Education

College of Engineering, University of California, Santa Barbara, CA 93106, United States

^* e-mail: pgunhal@ucsb.edu

Abstract

We present a suffix-aware pronunciation evaluation platform for Kannada language learning in diaspora educational contexts. Our system integrates speech-to-text transcription with novel scoring algorithms that target morphological accuracy, a challenge in agglutinative languages like Kannada. The system extracts five transcript-derived features, which are fed into a lightweight Multilayer Perceptron (MLP) regressor trained on 312 annotated recordings. Key contributions include a suffix-weighted loss function that penalizes morphophonemic deviations, the use of an Akshara parser to isolate morphological units, and a MERN platform providing feedback with sub-second latency. On validation data, the model achieved an RMSE of 0.121 and a Pearson correlation of r = 0.81 with instructor scores, outperforming traditional edit-distance and GOP baselines. Ablation studies and saliency analysis confirm that suffix-aware features boost model precision in identifying pronunciation errors. Our findings highlight the potential of suffix-aware NLP systems for regional language education and affirm the feasibility of AI-driven feedback for educational contexts.