This study investigates multimodal approaches for Speech Emotion Recognition (SER) in Bengali and English, using both audio and textual features. Initially, audio-only models like CNN, TabNet, Bagging with SVM, Random Forest, and XGBoost showed notable efficacy, with XGBoost achieving 88% accuracy. Hybrid models combining audio and text features—such as CNN with LSTM, XGBoost with Bi-LSTM, and CatBoost with Bi-LSTM—improved performance, with XGBoost with Bi-LSTM reaching 83% accuracy. On English datasets, BERT was the most effective, achieving 93% accuracy and outperforming Bi-LSTM with GloVe embeddings at 86%. Hybrid models with attention mechanisms, like CNN with LSTM, enhanced performance but did not surpass BERT. These findings highlight BERT’s superiority in detecting emotions in audio data and underscore the value of advanced language models in multilingual emotion recognition.