Towards End-to-End Synthetic Speech Detection

The constant Q transform (CQT) has been shown to be one of the most effective speech signal pre-transforms facilitate synthetic detection, followed by either hand-crafted (subband) cepstral coefficient (CQCC) feature extraction and a back-end binary classifier, or deep neural network (DNN) directly for further classification. Despite rich literature on such pipeline, we show in this paper that pre-transform features could simply replaced end-to-end DNNs. Specifically, experimentally verify only using standard components, light-weight outperform state-of-the-art methods ASVspoof2019 challenge. proposed model is termed Time-domain Synthetic Speech Detection Net (TSSDNet), having ResNet- Inception-style structures. We demonstrate models also have attractive generalization capability. Trained ASVspoof2019, they achieve promising detection performance when tested disjoint ASVspoof2015, significantly better than existing cross-dataset results. This reveals great potential DNNs without features.