Machine learning for automatic classification of Ge’ez noun reading styles in Ethiopic prosody

This study introduces a machine learning framework to classify the four traditional oral noun reading styles of Ge’ez: Seyaf ( ), Wedaki ( ), Tetay ( ), and Tenesh ( ). Using a manually curated corpus of 4264 nouns, we propose a linguistically constrained feature extraction method utilizing character n-grams (1–12) to capture sub-word phonological cues. To mitigate data scarcity and class imbalance, we implemented character-level augmentation and fold-specific SMOTE. In repeated stratified cross-validation, Random forest achieved 96.73 ± 0.14% accuracy, outperforming Linear SVC (95.64 ± 0.25%). A Wilcoxon signed-rank test confirmed a statistically significant difference (W = 0.0000, p = 5.96 × 10 −8 ). The accuracy difference (Δ = 0.58%, 95% CI [0.47%, 0.69%]) indicated a large effect size (paired Cohen’s d = 2.22). This research establishes a quantitative computational baseline for low-resource classical languages, demonstrating that complex oral prosody can be predicted solely through structured orthographic features, supporting digital heritage preservation.