Depression, an illness affecting more than 280 million people worldwide, manifests itself in daily anxiety and disrupts social relationships, which can lead to physical problems and even suicide. However, a large proportion of people suffering from depression choose not to seek medical help due to social stigma, cost, and lengthy detection procedures, highlighting the need for non-clinical diagnostic methods. Social media provide a valuable platform for the expression of emotions through messages. The analysis of these messages offers an interesting prospect for the early detection of depression, which could help prevent associated illnesses or suicide. However, detecting depression from text using natural language processing (NLP) remains a major challenge. This research proposes a comparative analysis of four different NLP models: CNN-BiLSTM with attention, BERT, RoBERTa and XLNet. We evaluate their performance in detecting depression using a Reddit dataset. The results reveal that BERT is the best model, with a recall of 97.52%.