Publicly available datasets analysis and spectrogram-ResNet41 based improved features extraction for audio spoof attack detection

The rapid expansion of voice-based technologies across diverse applications underscores the critical need for robust security measures against audio spoofing attacks. This paper comprehensively examines publicly available datasets that have been developed to detect audio spoof attacks. The research encompasses a compilation of datasets, including ASVspoof dataset series (2019, 2021), Voice Spoofing Detection Corpus (VSDC), Voice Impersonation Corpus in Hindi Language (VIHL) and DEepfake CROss-lingual evaluation dataset (DECRO), covering various spoofing attack scenarios of English, Hindi and Chinese languages. In the first part of the paper, a baseline for the proposed research work has been developed by comparing the performances of state-of-the-art baseline Linear frequency cepstral coefficient (LFCC) features with application of four different machine learning classifiers Random forest (RF), K-nearest neighbor (KNN), eXtreme gradient boosting (XGBoost), and Naïve Bayes (NB) at the backend, over these four different datasets. In second part of the proposal, we have used novel feature combination of Mel Spectrogram-Residual Network41 (ResNet41)-Linear discriminant analysis (LDA) and Gammatone Spectrogram-ResNet41-LDA, one by one, with application of same set of machine learning classifiers at the backend. The combination of Gammatone spectrogram-ResNet41-LDA along with XGBoost classifier has achieved an Equal Error Rate (EER) of 1.7, 1.28, 0.5, 0.36, 0.03, 0.07, and 0.9% for ASVspoof 2019 Logical Access (LA), ASVspoof 2019 Physical Access (PA), ASVspoof 2021 Deepfake, VSDC, DECRO English, DECRO Chinese, and VIHL datasets, respectively. Hence, the proposed research work in this paper achieves the objective of assessing the feasibility and utility of publicly available state of the art datasets for training and testing advanced algorithms in identifying manipulated audio.