Research on Acoustic Feature Extractor for Automatic Speaker Verification Systerm

In this paper, four frontends which are acoustic feature extractors are designed based on PNCCs extractor by removing its medium-time processing, power-law nonlinearity and DCT and by introducing per-channel energy normalization (PCEN), enhancing the robustness of ASV system. Of the four frontends, two frontends have learnable PCEN whose parameters are optimized automatically. These four acoustic feature extractors and MFCCs extractor are used as the frontends of ECAP A-TDNN respectively, resulting in five DNN-based speaker embedding extractors. Based on five speaker embedding extractors, five ASV systems are constructed respectively. All four ASV systems with designed frontends outperform the ASV system with MFCCs extractor in both intra-domain and cross-domain scenarios. On VoxCeleb-O evaluation dataset, the ASV system with learnable PCEN and mean power normalization, which is named as MLPs-ASV in the paper, achieves an improvement of 10.38% in EER, and an improvement of 11.86% in minDCF, respectively, compared to the ASV system with MFCCs extractor. On VoxMovies evaluation dataset, MLPs-ASV achieves 12.96% improvement in EER than the ASV system with MFCCs extractor.