Parkinson’s Disease (PD) detection using the biomarkers derived from speech signals has shown promising results in this field. At early stage of PD, the symptoms are subtle and autoencoder can learn the more abstract and high level features, while Transfer Learning (TL) can help the knowledge obtained from the autoencoder to train a Deep Neural Network (DNN). Hence, in this paper, a novel Transfer Learning based Autoencoder model has been proposed for the detection of PD. The proposed method is evaluated using a popular publicly available dataset: UCI Oxford Parkinson’s Disease Detection dataset and has compared with four benchmark models, namely Multilayer perception (MLP), Long-Short Term Memory (LSTM), standalone Autoencoder and Attentive Interpretable Tabular Learning (TabNet). The suggested method obtained an accuracy of 97.44%, precision of 1, recall of 0.968 and Gmean of 0.984 outperforming other models. The simulation results demonstrates that this hybrid model outperforms numerous state-of-the-art techniques. It can assist in developing an automated, reliable, and accurate PD detection model.