Automatic Modulation Classification (AMC) plays a crucial role in Cognitive Radio (CR) systems, especially within Internet of Things (IoT) devices where spectrum efficiency and flexibility are paramount. Traditional modulation classification methods often rely on feature extraction and machine learning (ML) algorithms, which need a lot of complex calculations and may struggle with complex modulation schemes and noisy channels. Deep Learning (DL), particularly Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs), has a positive impact on AMC due to their ability to automatically learn and extract discriminative features from the sequences of raw I/Q received signals. So, this paper proposes DL AMC model is built and investigated using Radioml2016 data for enhancing spectrum management for CR in IoT systems. The proposed model has reduce model parameters by 35% by using depthwise separable convolutional and traditional convolution to create the model architecture while increasing the accuracy of the model to 84 % at high SNR . The reduction in the model parameters led to a reduction in the prediction time to achieve the requirement in cognitive radio systems for IoT devices.