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

The paper introduces QAMO, a Quality-Aware Multi-Centroid One-Class Learning approach for detecting speech deepfakes, which enhances traditional one-class learning by incorporating multiple centroids that represent distinct quality subspaces of bona fide speech. By utilizing a multi-centroid ensemble scoring strategy, QAMO achieves an equal error rate of 5.09% on the In-the-Wild dataset, surpassing previous models. This method is significant for practitioners as it improves the detection of deepfakes by accounting for intra-class variability in speech quality, reducing reliance on quality labels during inference.