In recent years, detecting emotions has become an important open research area, with applications ranging from mental health support to customer services and entertainment. In this study, we have done a comparison of BERT model base configuration variants for emotion detection using two distinct benchmark datasets: MELD and GoEmotions. Our work proves the potency of fine-tuned transformer models, with BERT and RoBERTa models giving quite good results. However, the performance was not consistent for all the emotions, and thus, there is a need to fine-tune the hyperparameters. In particular, the analysis of the BERT model’s performance on the core emotions was informative and showed its strengths and weaknesses. This study contributes to the advancement of robust and effective emotion recognition models, with significant implications for various applications. The results stress the importance of the careful choice of the preprocessing pipeline and the model in emotion recognition tasks. More importantly, the best accuracy of 62% was obtained using RoBERTa on the MELD dataset and 65% using BERT on the GoEmotions dataset. Thus, our research focuses on the aspects of text that have been left untouched in other works, making it unique in this field.