This paper mainly presents our developed approach for the CNSRC2022 competition, specifically the open and fixed tracks in speaker verification task. In the context of speaker verification, a standard protocol is to extract the discriminative feature embeddings to determine the speaker identity via the similarity calculation. Compared to the VoxCeleb datasets, the CN-Celeb datasets involve more complex conditions as well as more challenging scenarios, which increases multi-genre and cross-genre complexity greatly. For fixed track, we have proposed two main improvement options. In terms of the model architecture, adaptive convolution extracts more robust representations, while dynamic convolution improves the representation capacity of the model. In terms of the task, we find that the noisy scene information could bring the negative effect. To handle this problem, we adopt a gradient reversal layer to decouple the harmful scene features. For open track, we use a pre-trained model trained on the VoxCeleb datasets, and then fine-tune it on the CN-Celeb datasets. Finally, by fusing the scores of each system, our method achieves 0.4195 minDCF in the fixed track and 0.3707 minDCF in the open track.
Cite as: Zhang, J., Ren, Q., Qin, Y., Wan, Z., Mao, Q. (2022) Cross-Scene Speaker Verification Based on Dynamic Convolution for the CNSRC 2022 Challenge. Proc. The Speaker and Language Recognition Workshop (Odyssey 2022), 368-375, doi: 10.21437/Odyssey.2022-51
@inproceedings{zhang22c_odyssey,
author={Jialin Zhang and Qinghua Ren and Youcai Qin and Zikai Wan and Qirong Mao},
title={{Cross-Scene Speaker Verification Based on Dynamic Convolution for the CNSRC 2022 Challenge}},
year=2022,
booktitle={Proc. The Speaker and Language Recognition Workshop (Odyssey 2022)},
pages={368--375},
doi={10.21437/Odyssey.2022-51}
}