Explainable AI for forensic speech authentication within cognitive and computational neuroscience.

Area of Science:

• Digital Forensics
• Artificial Intelligence
• Speech Processing

Background:

• Deepfake audio poses significant threats to forensic speech authentication.
• Existing detection methods struggle with the sophistication of manipulated audio.

Purpose of the Study:

• To develop a deep learning framework for enhanced detection of deepfake audio.
• To improve the accuracy and generalization of manipulated audio detection systems.

Main Methods:

• A hybrid deep learning model combining Convolutional Neural Networks (CNNs) for spectral feature extraction and Long Short-Term Memory (LSTM) networks for temporal modeling.
• Utilized Linear Frequency Cepstral Coefficients (LFCCs) as acoustic features.
• Applied explainable artificial intelligence (XAI) techniques like Grad-CAM and SHAP for model interpretability.

Main Results:

• The CNN-LSTM model achieved superior accuracy and generalization on ASVspoof2019 LA and WaveFake datasets.
• LFCC features outperformed Mel-Frequency Cepstral Coefficients (MFCC) and Gammatone Frequency Cepstral Coefficients (GFCC).
• XAI analysis revealed the model's focus on high-frequency artifacts and temporal inconsistencies, validating its forensic relevance.

Conclusions:

• The proposed framework offers a robust and interpretable solution for forensic audio authentication.
• The integration of XAI enhances trust and transparency in deepfake detection.
• LFCC features demonstrate significant forensic value in identifying manipulated audio.