The Cognitive Radio (CR) is recommended to improve the utilization of wireless spectrum, and essential phase of CR is spectrum sensing. The traditional spectrum sensing methods relied on detection of energy and extraction of features from the received signal at particular location. In this research, the proposed Covariance matrix—Long Short-Term Memory (Co-LSTM) method for classifying the signals in CR networks. The Co-LSTM method included the weight metrics for enhancing the performance of LSTM method for signal classification. The dataset is generated for classifying the signals in the research and it is classified by using the Co-LSTM method. The Co-LSTM method effectively classified the signals in the CR networks with high accuracy and less error rate. The Co-LSTM method is analyzed with metrics of accuracy, F1-score, Mean Square Error (MSE), and Root MSE (RMSE). The Co-LSTM method attained 98.42%, F1-score 97.69%, MSE 0.015, and RMSE 0.132 which is superior than existing methods like Convolutional Neural Network (CNN), Bidirectional LSTM (Bi-LSTM), and Transformer Networks (TN).