Software defects increase development cost, delay release schedules, and reduce user satisfaction. Conventional defect prediction methods often rely on rigid assumptions, linear relationships, and manually selected thresholds. Soft computing provides an alternative by modeling uncertainty, imprecision, and nonlinear relationships in software engineering data. This paper reviews the application of fuzzy logic, artificial neural networks, evolutionary computation, support vector machines, swarm intelligence, and hybrid methods to software defect prediction and quality improvement. The study analyzes the role of software metrics, data preprocessing, feature selection, model training, and performance evaluation. A conceptual framework is proposed in which repository data and process metrics are transformed into defect-risk estimates and quality-improvement recommendations. The analysis indicates that hybrid soft computing models can improve prediction capability when compared with individual techniques, particularly in the presence of class imbalance and noisy project data. However, issues related to interpretability, data quality, cross-project generalization, and model maintenance remain significant. The paper concludes that soft computing should be integrated with software quality assurance processes rather than treated as an isolated prediction mechanism.