On the Generalization of Fused Systems in Voice Presentation Attack Detection

This paper describes presentation attack detection systems developed for the Automatic Speaker Verification Spoofing and Countermeasures Challenge (ASVspoof 2017). The submitted systems, using calibration and score fusion techniques, combine different sub-systems (up to 18), which are based on eight state of the art features and rely on Gaussian mixture models and feed-forward neural network classifiers. The systems achieved the top five performances in the competition. We present the proposed systems and analyze the calibration and fusion strategies employed. To assess the systems' generalization capacity, we evaluated it on an unrelated larger database recorded in Portuguese language, which is different from the English language used in the competition. These extended evaluation results show that the fusion-based system, although successful in the scope of the evaluation, lacks the ability to accurately discriminate genuine data from attacks in unknown conditions, which raises the question on how to assess the generalization ability of attack detection systems in practical application scenarios.