Multi-Scene Robust Speaker Verification System Built on Improved ECAPA-TDNN

Abstract

In order to solve the problems of cross-domain, short speech, and noise interference in industrial application scenarios of speaker recognition, this paper proposes an improved ECAPA-TDNN for a multi-scene robust speaker verification system architecture-improved DD-ECAP A-TDNN.The design of the DD-ECAPA-TDNN architecture is inspired by the model ECAPA-TDNN, which has recently become popular in ASV systems. Firstly, we use FBanks to extract acoustic features, followed by the DD-SE-Res2Net Block proposed in this paper to capture local features efficiently. Finally, the output feature mapping of all DD-SE-Res2Net Blocks aggregated at multiple scales, and finally the ASP pooling operation is performed. The experiments were based on the VoxCeleb1-dev dataset, and SC-AAMSoftmax was used to train a speaker identification model for 1211 speakers. This DD-ECAPA-TDNN model was used as speaker embedding extractor to construct an automatic speaker verification (ASV) system. We used VoxMovies and VoxCeleb1-O evaluation sets to simulate three scenarios of cross-domain, short speech and noise interference, respectively, to evaluate the performance of the DD-ECAPA-TDNN system under multiple scenarios. The system achieves an EER of 2.51% on VoxCeleb1-O. The DD-ECAPA-TDNN system significantly outperforms the ECAPA-TDNN system in terms of recognition performance in multiple scenarios. Finally, our ablation experiments show that the DD-SE-Res2N et Block has a positive impact on the performance of the ASV system, as well as that the DD-ECAPA-TDNN can extract robust and accurate speaker embedding with good scene generalization.