The disease called Parkinson’s when progresses, it causes nerve cells to die. The symptoms can be motor and non-motor types and its early detection is an important concern. The study suggests a new way for deep learning to work using a Long Short-Term Memory (LSTM) along with a 3D-CNN. The training process for these (LSTM) layers involves a Genetic Algorithm (GA) when selecting features, to use the given voice signal data to diagnose PD dataset. The database consists of 195 voice recordings that each have 23 features. Measured values such as fundamental frequency, jitter, shimmer and nonlinear parameters. Spectrograms of voice signals are used as input in our hybrid model. Identifies how events are distributed over space and time and perform well Instead of using traditional machine learning or single deep learning models. The results of experiments show an accuracy percentage of 94.8% and an accuracy of 95.2% and a specificity of 93.5% are achieved which is higher than baseline results provided by Support the methods used are SVMs and two-dimensional Convolutional Neural Networks (2D-CNNs). It draws attention to the ways technology could be used by using advanced deep learning, it is setting the stage for machines that are small, easy to use and suitable for many people.