This study evaluates the performance of BERT, AlBERT, roBERTa, distilBERT, and DeBERTa models in diagnosing early Alzheimer’s Disease (AD) based on scripts documented from spoken speech. The classification leverages natural language techniques, transfer learning, and object-oriented programming (OOP). The study leveraged the ADReSS dataset from DementiaBank, including 156 speech samples and associated transcripts from non-AD (N = 78) and AD (N = 78). Participants described the Cookie Theft picture, and their speech was recorded and transcribed. We exclude non-natural components and pad (add tokens) or truncate (cut tokens) to format the dataset. Then, we applied transfer learning with pre-trained Transformer models—BERT, AlBERT, roBERTa, distilBERT, and DeBERTa—and evaluated the results using precision, recall, and Fl-score metrics. The BERT model achieved an accuracy of 73% and an F1 score of 79% using 30% of samples from the DementiaBank dataset for evaluation. The fine-tuned DeBERTa model achieved an accuracy of 72% with an F1 score of 58%. Both BERT and DistilBERT outperformed the fine-tuned ALBERT model which had an accuracy of 77% with an F1 score of 62%. The fine-tuned DistilBERT model demonstrated the best performance with an accuracy of 77% and an F1 score of 79%. The fine-tuned RoBERTa model performed the least favorably, with an accuracy of 69% and an F1 score of 62%. Moreover, after implementing the OOP approach, the highest accuracy observed of the BERT model in predicting AD is 83%, equivalent to accurately predicting 20 out of 24 cases. This study extensively assesses BERT, AlBERT, roBERTa, distilBERT, and DeBERTa models in early AD diagnosis via speech transcripts. It highlights the potential real-life application of AD diagnosis based on speech, emphasizing its accessibility compared to traditional AD diagnosis methods.