Texture Analysis of Edge Mapped Audio Spectrogram for Spoofing Attack Detection

The vulnerability of the Automatic Speaker verification technology to different presentation attacks has drawn much interest to design anti-spoofing countermeasures. Following their success in face anti-spoofing, some techniques based on texture analysis were cross-pollinated recently to the voice anti-spoofing domain with very promising results. In this same research direction and motivated by the fact that the spectrogram image local gradient amplitude is sensitive to image distortions presented by spoof attacks, a novel voice anti-spoofing countermeasure based on texture analysis of spectrogram image is devised in this paper. The main novelty of the proposed approach resides in the use of edge detection techniques applied to the spectrogram image. The recorded speech was initially converted to a 2-D image. Then, the spectrogram is mapped by the use of the edge detection method. Finally, different texture descriptors are used to extract the visual features from the original and the mapped spectrograms combination. Experimental results on the ASVspoof 2017 v2.0 dataset show that the proposed countermeasure clearly outperforms some well-known techniques, providing an Equal-Error Rate (EER) as low as 3.52% and 22.86% for the development and evaluation sets, respectively. For the ASVspoof 2019 dataset, in turn, the proposed method achieves up to 27.62% and 29.15% relative EER improvement over the baseline CQCC feature in the Logical and Physical Access scenarios, respectively. In addition, the results show that the proposed hand-crafted features-based approach with an SVM classifier can deliver superior performance than more sophisticated solutions that rely upon the complex neural network architecture.