Parkinson’s disease is a neurodegenerative condition caused due to the lack of secretion of dopamine hormone in the brain. Motor and non-motor activities get hamper in Parkinsonism. Detection of Parkinson's disease (PD) in the early stage using voice signals has been used as a promising biomarker. This work proposed a unique feature extraction method from spectrogram of voice combining the convolutional neural network (CNN) and bottleneck encoder architecture to detect PD. The flatten layer output of CNN is fed to the bottleneck autoencoder layer and the output of autoencoder is used as a feature to represent the voice signal. The performance of the proposed method is assessed with the Italian language dataset of sustained vowels comprising 22 healthy and 28 PD speakers. The Support Vector Machine (SVM) and XGBoost classifiers are used on the extracted features. The SVM classifier provided the highest average accuracy of 0.96 in the vowel /u/. The experimental results showed the potential of the proposed method for precise and non-invasive PD detection using voice signals.