An Analytical Approach to Punjabi Dialect Identification using Feature Selection and Ensemble Models

In this work, Punjabi dialect identification system is developed by first preparing dialectal speech database and then using eight prosodic and 10 spectral shape features for capturing dialect-specific variations. To enhance classification performance, three ensemble learning techniques—Boosting, Bagging, and Stacking—are employed. Four feature selection algorithms, namely, Least Absolute Shrinkage and Selection Operator (LASSO), Recursive Feature Elimination (RFE), Analysis of Variance (ANOVA), and Principal Component Analysis (PCA) are implemented and performances are compared for various combinations of speech features. Accuracy, precision, recall, and F1 scores are the performance metrics used to evaluate the performance of the system. The best performance, both in terms of accuracy and F1 score, falls in range of 0.80–0.82, when combination of prosodic and spectral shape features is taken for All_Feature_Set and LASSO_Set using Boosting model. In comparison, Stacking model achieves moderately strong results, with accuracy and F1 scores between 0.75–0.79 for the same feature sets. The Bagging model also performs well, though its outcomes remain slightly below those of Boosting and Stacking. The RFE_Set also delivers strong results, closely matching the performance of the LASSO_Set. In contrast, the ANOVA_Set and PCA_Set show comparatively weaker performance across all models. The results are tested for 80:20 split and 5-fold, 10-fold cross validations which provided nearly similar outcomes. So, findings of this work conclude that prosodic features in comparison to spectral shape features carry less dialectal information but the combination of these features performed exceptionally well.