Cyberbullying is a pressing issue in the digital age, especially in non-English-speaking communities. This study focuses on using deep learning techniques to detect cyberbullying in the Bangla language. The research employs a Bangla dataset that includes both cyberbullying and non-cyberbullying text, which undergoes preprocessing, tokenization, and sequence transformation for input into a LSTM recurrent neural network. The LSTM model incorporates word embedding and dropout techniques to enhance its performance. Evaluation metrics like accuracy and AUC-ROC score are employed to assess the model's effectiveness. The study investigates many methods, such as vernacular-specific preparation and data augmenting, that are specifically designed for the identification of cyberbullying in the Bangla language. The findings demonstrate deep learning's capacity to identify cyberbullying in Bangla, which makes a significant contribution to resolving this problem in non-English settings. The research reveals impressive accuracy rates, with LSTM achieving 99.99% accuracy, and a different approach, Convolutional Neural Network (CNN), yielding 99.022%. Ensemble models, including Gradient Boosting, XGBoost, and Random Forest, are also employed and optimized using hyperparameter tuning. Notably, LSTM stands out as the most reliable classifier for cyberbullying detection, as confirmed by the experimental findings and recent research.