The importance of software defect prediction in ensuring software quality, reducing development costs, and enhancing efficiency through the initial identification of potential defects in the development process is of utmost significance. Recent developments in deep learning have made it possible to automatically extract complicated characteristics from software data, significantly enhancing SDP. In order to create a stacked-ensemble model for SDP, this study uses DL models, more especially Convolutional Neural Networks, Long Short-Term Memory networks, and Gated Recurrent Units. Additionally, Optuna, an automated hyperparameter tuning framework, is utilized to optimize the DL models. The experiments conducted demonstrate significant improvements in predictive performance, with CNNs achieving an average enhancement of 20.87%, LSTMs 19.75%, and GRUs 15.05%. The proposed ensemble model exhibited the highest overall performance, ranking first in our evaluations, with Friedman’s test confirming the statistical significance of the results. These results emphasize the significance of hyperparameter tuning in SDP and demonstrate the potential of optimized DL techniques in enhancing software defect prediction, offering a robust and reliable approach for maintaining the quality of software.