$\beta$-Order Energy-Weighting Modulation on Spectral Bins for Replay Speech Detection

The usage of recording and playback devices for the generation of replay speech introduces some variation into replay speech. This kind of variation penetrates into the shorttime details of the replay speech signal in a way of energy distribution. Together with the environmental noise during the process of recording, the disparity between replay speech and genuine speech is discernible. In this paper, in order to emphasize the disparity of the short-time energy distribution between replay speech and genuine speech, we propose a β-order energyweighting (EW) method to modulate the spectral-energy into spectral bins in short time frame-wise duration. Specifically, we apply the β-order EW modulation in different spectral domains, i.e., octave spectral domain, Mel spectral domain and linear spectral domain. Subsequently, we propose three novel features: constant-Q EW octave coefficient (CEOC), Mel frequency EW cepstral coefficient (MFECC), and linear frequency EW cepstral coefficient (LFECC). Moreover, with the three proposed features, relevant replay speech detection (RSD) systems are developed with respective ResNet and DNN as backend classifiers. With experiments conducted on respective ASVspoof 2017 version 2.0, BTAS 2016 physical access, ASVspoof 2019 physical access and ASVspoof 2021 physical access corpora, the effectiveness of the three proposed features is evaluated in terms of the RSD performance. The evaluation results demonstrate that the β-order EW modulation exhibits superior discriminative capability in detecting replay speech. Furthermore, through the experiments, it can be observed that our proposed RSD systems outperform many state-of-the-art RSD systems.