iWAX: interpretable Wav2vec-AASIST-XGBoost framework for voice spoofing detection.

Area of Science:

• Speech processing
• Machine learning
• Artificial intelligence

Background:

• Deep learning models like wav2vec 2.0 (w2v2) are increasingly used for voice spoofing detection.
• However, their complex nature hinders interpretability, making it difficult to understand their decision-making process.

Purpose of the Study:

• To develop an interpretable voice spoofing countermeasure (iWAX) that combines w2v2 with XGBoost.
• To enable explanations of detection predictions by identifying important temporal and frequency segments.

Main Methods:

• iWAX utilizes a fine-tuned w2v2 front-end and AASIST back-end with an XGBoost classifier.
• Sinc filters are applied for frequency band analysis, and temporal analysis focuses on key w2v2 features.
• XGBoost's feature importance mechanism is central to the interpretability approach.

Main Results:

• iWAX demonstrated superior performance compared to baseline models (AASIST, w2v2-AASIST) on the ASVspoof 2019 LA dataset.
• The system provided human-understandable explanations for its voice spoofing detection predictions.
• Robustness was confirmed using LightGBM, indicating broad applicability of the interpretability method.

Conclusions:

• iWAX achieves a strong balance between high performance and interpretability in voice spoofing detection.
• The approach addresses the limitations of traditional and deep learning-based countermeasures.
• This work facilitates trust and understanding in AI-driven audio security systems.