Answers are grounded in this paper and cite their sources.
Audio deepfake detectors often exhibit substantial performance degradation when evaluated on speech generated by synthesis technologies not represented during training. In this study, 20 individual detectors, derived from five deep learning architectures (DenseNet121, ResNet50, InceptionV3, ResNet34, and VGG16) and four spectral representations (LOG, Mel, DWT, and CQT), achieved F1-scores above 0.95 during internal validation but exhibited F1-macro values as low as 0.3460 on external audio, highlighting the challenges of generalization. To address this problem, we propose Data-Centric and Model-Centric Fusion (DCMC-Fusion), an ensemble learning framework that jointly investigates the composition and size of the ensemble and the classifier used to combine model predictions. The outputs of the 20 individual models are combined into 31 datasets with different architectural compositions and ensemble sizes. Four supervised fusion classifiers (Logistic Regression, Support Vector Machine, Random Forest, and XGBoost) are evaluated, resulting in 124 configurations. Experiments involving 2484 Spanish-language audio samples, equally distributed between natural and synthetic speech, show that DCMC-Fusion achieves a maximum F1-macro of 0.7184 using three architectures and XGBoost. This represents an improvement of 4.33 percentage points over the best individual detector (0.6751). Furthermore, DCMC-Fusion consistently outperforms One-Hard Voting, Soft Voting, and Weighted Averaging across all ensemble sizes. These findings demonstrate the benefits of jointly investigating ensemble composition and fusion mechanisms for audio deepfake detection under challenging external evaluation conditions.
No takes yet. Share an insight, caveat, or question.
Ballesteros et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: