Parkinson disease (PD) being a progressive neurological condition that affects the motor control which include vocal analysis, speech production for early non-intrusive detection. In this work, speaker-indepen-dent framework is automated to detect the speech by utilizing hybrid feature set produced by incorporating OpenSMILE features and nonlinear features like Hjorth parameters, entropy measures and fractal dimensions. Machine learning models—Support Vector Machine, Random Forest, K-Nearest Neighbor and Naive Bayes along with Deep learning model Convolutional Neural Network with Bidirectional Long Short-term Memory (CNN+BilSTM) is been trained and evaluated by employing the hybrid features. To maintain generalizability to unseen speakers, speaker-independent train-test split is adopted and SMOTE addresses the class imbalance through over sampling. Through comparative analysis, it reveals that hybrid feature set improves the classification performance when compared to linear and non-linear executed independently. Random Forest secured the highest accuracy and among the traditional Machine learning models and Deep learning models. By integrating complementary feature representation, we obtain a robust and reliable framework for early PD detection with major applications in continuous patient monitoring.