hustep-lab/VoxVietnam-Dataset

The dataset contains three subsets:

• train: Official training set (VoxVietnam-T) used in the paper (1,256 speakers, 161,457 samples).
• train_small: VoxVietnam-T-small, sampled from VoxVietnam-T to have the same size as Vietnam-Celeb (879 speakers, 83,000 samples).
• The VoxVietnam-T-noisy in the paper is not uploaded since it is not clean for supervised training, just for ablation studies in the paper only.

[Update 29 Mar, 2025] The VoxVietnam-E and VoxVietnam-H are labelled by volunteers without visual information. Our team released another independent test set, called VoxVietnam-O, verified by us by listening and watching the video segments for the highest accuracy. The speakers in VoxVietnam-O are sampled from the test partition. You can download the data and test list for VoxVietnam-O here. We encourage researchers to use VoxVietnam-O for evaluation.
Here are the results on VoxVietnam-O for reference. We use Ruijie Tao's implementation of ECAPA-TDNN:

Train	EER (%)	minDCF (%)
VoxVietnam-T	3.03	0.4781
Vietnam-Celeb-T	3.25	0.5376
VoxVietnam-T-small	3.96	0.5273
VoxVietnam-T-noisy	6.91	0.6813
Vietnam-Celeb-T + VoxVietnam-T	3.34	0.5286

[Update 03 Jan, 2025] Our paper has been accepted to ICASSP 2025! The preprint is available at: https://arxiv.org/abs/2501.00328.

Please cite our work as:

@INPROCEEDINGS{10890124,
  author={Vu, Hoang Long and Dat, Phuong Tuan and Nhi, Pham Thao and Hao, Nguyen Song and Thu Trang, Nguyen Thi},
  booktitle={ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={VoxVietnam: a Large-Scale Multi-Genre Dataset for Vietnamese Speaker Recognition}, 
  year={2025},
  volume={},
  number={},
  pages={1-5},
  doi={10.1109/ICASSP49660.2025.10890124}}