Automatic classification of speakers based on regional origin has applications in voice authentication, human–computer interaction, and sociolinguistic studies. This work investigates district-level voice classification for Bangla speakers from Dhaka and Chittagong. A dataset of 1,526 voice recordings was collected, and class imbalance was addressed using the RandomOverSampler algorithm, yielding 1,870 balanced samples. Mel-Frequency Cepstral Coefficients (MFCCs): MFCC was extracted and used to represent the spectral characteristics of speech. Six machine learning model i.e. Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), XG Boost and Gradient Boosting models were tested on this dataset. Experimental results show Logistic Regression algorithm gave the highest accuracy (99.47%), SVM algorithm accuracy is 99.20% and Random Forest algorithm accuracy is 98.40%, which indicates that MFCC features are highly discriminative and linear separable in the point of classification the district. These results set a baseline for regional voice recognition in Bangla and show that relatively simple models can be used to get excellent performance, as a basis for more advanced work in dialect recognition, accent analysis, and broader sociolinguistic applications.