This study presents a method for multi-class classification of crime-related Bengali newspaper headlines using ensemble machine learning techniques. The methodology begins with the development of a novel dataset categorized into eight distinct crime classes. Preprocessing involves TF-IDF-based bigram feature extraction to retain contextual meaning, followed by one-hot encoding of labels to ensure compatibility with machine learning models. Several base classifiers, including Logistic Regression, Random Forest, Naïve Bayes, and Support Vector Machine, were individually fine-tuned using Grid Search with cross-validation. A soft-voting ensemble model was then constructed, integrating the weighted outputs of the top-performing classifiers. The proposed ensemble approach achieved strong classification performance, demonstrating accuracy of 94%, precision of 94%, recall of 94%, and ROC AUC of 99%. Comparative analysis with existing methods confirms the superiority of this model in handling Bengali crime headline classification. This research contributes to the advancement of intelligent news analysis in Bengali and supports future developments in interpretable NLP applications for crime monitoring and public safety enhancement.