Smartphone Audio Replay Attacks Dataset

Smartphone based biometric applications are increasing exponentially in recent years. The challenges due to presentation attacks in biometrics have emerged to cause a potential vulnerability, limiting the reliability of biometrics for secure applications. In speaker recognition, audio replay attacks have demonstrated a severe threat to automatic speaker verification (ASV) systems. Alongside, the difference in language for enrolment and testing has displayed some impact on speaker recognition. In this direction, we have created a novel audio replay attack dataset for four different languages using smartphones as playback and recording devices. We have collected data in two different scenarios where the attack recording sensor and bona fide sensor are the same and different. The captured dataset is used for testing the vulnerability on both state-of-the-art speaker recognition method and commercial-off-the-shelf (COTS) method from VeriSpeak. The baseline presentation attack detection methods are benchmarked on replay attacks in a cross-smartphone scenario. The results show that the replay attacks indicate a severe threat towards the ASV methods, especially in the cross-smartphone scenario.