In this paper, we present the IMU&Elevoc systems submitted to the Short-duration Verification Challenge (SdSVC) 2021. Our submissions focus on both text-dependent speaker verification (Task 1) and text-independent speaker verification (Task 2). First, we investigate several frame-level feature extractor architectures based on ResNet, Res2Net and TDNN. Then, we integrate Squeeze-Excitation block and dimension cardinality to further improve the Res2Net-based backbone network. In particular, we probe an effective transfer learning strategy that overcomes the lack of Task 1 datasets and improves in-domain performance. A knowledge distillation method fusing multiple models is proposed to obtain a stronger single model. Experimental results on the SdSVC 2021 show that our primary system yields 0.0500MinDCF in Task 1 (ranked as 4th) and 0.0448 MinDCF in Task 2 (ranked as 6th).
Cite as: Zhang, P., Hu, P., Zhang, X. (2021) Investigation of IMU&Elevoc Submission for the Short-Duration Speaker Verification Challenge 2021. Proc. Interspeech 2021, 2322-2326, doi: 10.21437/Interspeech.2021-743
@inproceedings{zhang21o_interspeech,
author={Peng Zhang and Peng Hu and Xueliang Zhang},
title={{Investigation of IMU&Elevoc Submission for the Short-Duration Speaker Verification Challenge 2021}},
year=2021,
booktitle={Proc. Interspeech 2021},
pages={2322--2326},
doi={10.21437/Interspeech.2021-743}
}