Emotion extraction from text is a difficult but important task in natural language processing (NLP), especially in applications such as sentiment analysis, mental illness diagnosis, and human-computer interaction. This paper is a comparative performance study of deep learning (DL) models viz. Bi-LSTM, Bi-GRU, ANN, CNN, BERT, RoBERTa, and GPT on three benchmark corpora: ISEAR, GoEmotion, and MELD. Each of the datasets, varying in size and level of granularity of sentiment labels, was preprocessed uniformly using text cleaning, normalization, tokenization, and encoding to make them uniform and model-ready. We trained and fine-tuned each model with equal training parameters and tested them using standard metrics: F1-score and accuracy. Results indicate that Transformer-based models (RoBERTa and BERT) perform better than baseline models in all the datasets with superior performance to identify contextualized nuances of text data. The findings indicate real-world applications of pre-trained language models to achieve reliable emotion detection, with additional references to computational limitations and dataset imbalance that require further study.