Synthesized Speech Detection Based on Spectrogram and Convolutional Neural Networks

The paper presents systems based on convolutional neural networks designed to classify genuine and artificially generated speech signals, which were evaluated on database for logical access designed for 3rd Automatic Speaker Verification Spoofing and Counter-measures Challenge (ASVspoof 2019). Proposed systems achieved remarkable results on the development set, but rather modest on the evaluation set, i.e. equal error rate on development set is 0 % and on evaluation set 9.57 %.