Real-World Speech Recovery Under Multiple Distortions: A Two-Stage Framework With Feature Consistency and Adversarial Fine-Tuning

Recent advances in speech restoration and bandwidth extension have significantly improved audio quality under controlled conditions. However, their performance often degrades drastically in real-world scenarios where severe distortions and bandwidth limitations coexist. These challenges become particularly critical under extreme conditions, such as low signal-to-noise ratios and heavily degraded spectral features, where conventional methods struggle to recover intelligible speech. Compound distortions severely disrupt the time-frequency structure of speech signals, undermining the robustness and perceptual quality of existing enhancement systems. Although generative approaches, such as generative adversarial networks, excel in producing perceptually plausible outputs, they remain highly sensitive to input instability, limiting their reliability in such conditions. To address these challenges, we propose a novel two-stage framework that leverages a robust pre-training strategy to jointly restore multiple speech distortions for quality improvement. The method integrates a Scale-Shift Linear Transformation, which effectively suppresses non-stationary noise while preserving essential speech components. Additionally, a GAN-based fine-tuning strategy is introduced to enhance perceptual realism and reconstruct coherent high-frequency details. Experimental results demonstrate that our model, trained solely on the VCTK corpus, achieves high-fidelity restoration across various distortion scenarios. Its strong generalization ability under unseen and challenging conditions highlights the practicality and robustness of the proposed method for real-world speech restoration applications.