TRACE: Training-Free Partial Audio Deepfake Detection via Embedding Trajectory Analysis of Speech Foundation Models

1. 1Utilize pre-trained speech foundation models to generate embeddings for audio segments
2. 2Analyze the embedding trajectory to identify inconsistencies indicative of deepfakes
3. 3Apply the TRACE method to detect partial audio deepfakes without requiring frame-level annotations or retraining

AI engineers and researchers working on audio deepfake detection can benefit from this approach as it eliminates the need for supervised training and can adapt to new generative models

💡 Speech foundation models can implicitly encode forensic features for deepfake detection

🔊 TRACE detects partial audio deepfakes without training! 🤖