QAMO: Quality-aware Multi-centroid One-class Learning For Speech Deepfake Detection

Researchers introduce QAMO, a machine learning system that improves speech deepfake detection by using multiple quality-aware centroids instead of a single centroid to model genuine speech. The approach achieves a 5.09% error rate on challenging real-world datasets, advancing security in voice authentication and synthetic media detection.

The emergence of sophisticated deepfake technology has created significant vulnerabilities in voice-based authentication systems and digital trust infrastructure. QAMO addresses a fundamental limitation in existing one-class learning approaches, which treat all legitimate speech as a uniform distribution around a single point. By incorporating speech quality metrics derived from Mean Opinion Score assessments, the system creates distinct models for high and low-quality genuine speech, capturing the natural variation in real human voice patterns. This multi-centroid strategy reflects a broader shift in machine learning toward more nuanced, contextual models that account for real-world variability rather than oversimplified assumptions. The practical implications extend beyond academic research into commercial applications where voice remains a critical authentication factor—from banking systems to voice assistants. Financial institutions and technology companies face escalating risks from synthetic voice attacks that can bypass existing security measures. QAMO's improved detection accuracy reduces false positives and negatives that plague current systems, potentially protecting billions in assets and preventing fraud at scale. The ensemble scoring approach also reduces dependency on quality labels during deployment, making the system more practical for resource-constrained environments. As deepfake quality continues improving, the arms race between detection and generation intensifies. This research represents meaningful progress in the detection side, though widespread adoption requires integration with existing voice biometric platforms and validation across diverse languages and acoustic conditions.

• →Multi-centroid architecture outperforms single-centroid models by capturing natural variation in genuine speech quality.
• →Achieves 5.09% equal error rate on in-the-wild datasets, significantly improving detection reliability.
• →Quality-aware approach reduces false positives and negatives critical for real-world deployment.
• →Ensemble scoring strategy eliminates need for quality labels during inference, improving practical deployment.
• →Addresses growing security vulnerabilities in voice authentication systems threatened by advancing deepfake technology.