Bangla Dialect Classification Across 11 Regions Using Machine Learning and Transformer-Based Models

Dialect identification for low-resource languages, such as Bengali, presents a significant challenge due to high linguistic similarity between dialects and pervasive class imbalance in available datasets. This paper establishes the first comprehensive performance benchmark for the BanglaDial corpus, a largescale dataset for this task. A comparative analysis of classical machine learning algorithms—including LinearSVC, Random Forest, Decision Tree and a Transformer-based BanglaBERT model is presented. A key contribution is a classical pipeline combining a LinearSVC classifier with a FeatureUnion of hybrid word and character TF-IDF n-grams. The analysis reveals that this highly-optimized classical pipeline achieves a new State-of-the-Art (SOTA) accuracy of 82.03%, demonstrating its feature set is essential for capturing both semantic and morphological dialectal markers. Importantly, this highly-optimized classical SVM significantly outperforms the baseline-fine-tuned BanglaBERT model. This result suggests that for this specific task, a robustly-engineered and optimized high-dimensional feature set is more effective than a baseline-fine-tuned contextual embedding model, establishing a new SOTA benchmark for 11 regional dialects classification.