Сравнительный анализ XAI-подходов для обнаружения синтетической речи

To date, speech synthesis systems have advanced to the point where the automatic detection of synthetic and deepfake speech has become increasingly challenging. At the same time, most detection models remain largely opaque “black boxes”, which limits their applicability in digital forensics, biometrics, and other critical systems. This study reviews and comparatively analyzes explainable artificial intelligence (XAI) methods applied specifically to synthetic and deepfake speech detection. Gradient-based, relevance-based, and model-agnostic approaches are examined, including Integrated Gradients, Deep Taylor Decomposition, Layer-wise Relevance Propagation, LIME, SHAP, and Grad-CAM. Particular attention is paid to the principles underlying explanation generation and the interpretability of explanations in relation to the acoustic characteristics of speech signals. Approaches to evaluating explanation quality are discussed, including faithfulness, stability, and sensitivity to perturbations, and the limitations of published studies are analyzed. The comparison provides a structured overview of the strengths and limitations of different XAI methods, identifies challenges in comparing existing results, and outlines promising directions for future research.