Attention Network with GMM Based Feature for ASV Spoofing Detection

Automatic Speaker Verification (ASV) is widely used for its convenience, but is vulnerable to spoofing attack. The 2-class Gaussian Mixture Model classifier for genuine and spoofed speech is usually used as the baseline in ASVspoof challenge. The GMM accumulates the scores on all frames in a speech independently, and does not consider its context. We propose the self-attention network spoofing detection model whose input is the log-probabilities of the speech frames on the GMM components. The model relies on the self-attention mechanism which directly draws the global dependencies of the inputs. The model considers not only the score distribution on GMM components, but also the relationship of frames. And the pooling layer is used to capture long-term characteristics for detection. We also proposed the two-path attention network, which is based on two GMMs trained on genuine and spoofed speech respectively. Experiments on the ASVspoof 2019 challenge logical and physical access scenarios show that the proposed models can improve performance greatly compared with the baseline systems. LFCC feature is more suitable for our models than CQCC in experiments.