EMBNet: Multi-scale feature learning with efficient channel attention for deepfake speech detection

Area of Science:

• Artificial Intelligence
• Digital Forensics
• Audio Signal Processing

Background:

• Generative AI advancements pose threats to digital audio authenticity.
• Deepfake speech presents challenges for forensic analysis and legal applications.
• Existing methods struggle with subtle acoustic anomalies in manipulated audio.

Purpose of the Study:

• To propose EMBNet, a task-oriented framework for deepfake speech detection.
• To enhance the representation of subtle acoustic anomalies for improved detection.
• To advance the accuracy of forensic audio authenticity assessment.

Main Methods:

• Integration of Efficient Channel Attention (ECA) for adaptive feature emphasis.
• Utilizing a multi-scale bottleneck to capture local and hierarchical spoofing traces.
• Balancing fine-grained local detail with global hierarchical representation.

Main Results:

• EMBNet significantly outperforms baseline and state-of-the-art methods on the ASVspoof 2019 dataset.
• Achieved an Equal Error Rate (EER) of 2.67%, Area Under the Curve (AUC) of 97.32%, and F1-score of 97.28%.
• Ablation studies confirm the effectiveness of ECA and multi-scale modules.

Conclusions:

• EMBNet demonstrates promising performance for forensic audio analysis.
• The framework improves the discrimination between genuine and manipulated speech.
• EMBNet supports enhanced forensic audio authenticity assessment.