Overview

This code is for "Segment Aggregation for short utterances speaker verification using raw waveforms". This paper proposed a method that compensates for the performance degradation of speaker verification, referred to as "segment aggregation". The proposed method adopts an ensemble-based design to improve the stability and accuracy of speaker verification systems.

**We referenced the baseline system RawNet code at here

Email kimho1wq@naver.com

Datasets

We used VoxCeleb2 dataset for training and VoxCeleb1 original evaluation set for test. Input two dataset in DB directory for training and test.

Training

1. ./train_run.sh
2. python train_RawNet_SA_TS_rand.py -name rawnet_SA_TS_rand_1s_3s

Test

Go into test directory
1. ./test_run.sh
2. python test_pre_trained_model.py -pretrained_name best_rawnet_SA_TS_1s_3s.pt

BibTex

This reposity provides the code for reproducing below papers.

@article{seung2020segment,
  title={Segment Aggregation for short utterances speaker verification using raw waveforms},
  author={Seung-bin Kim and Jee-weon Jung and Hye-jin Shim and Ju-ho Kim and Ha-Jin Yu},
  journal={arXiv preprint arXiv:2005.03329},
  year={2020}
}

Log

• 2020.05.08. : Init
• 2021.03.18. : issue bug fix