Introduction
This repository contains several materials that supplements the Spoofing-Aware Speaker Verification (SASV) Challenge 2022 including:
- calculating metrics;
- extracting speaker/spoofing embeddings from pre-trained models;
- training/evaluating Baseline2 in the evaluation plan.
More information can be found in the webpage and the evaluation plan
Prerequisites
Load ECAPA-TDNN & AASIST repositories
git submodule init
git submodule update
Install requirements
pip install -r requirements.txt
Data preparation
The ASVspoof2019 LA dataset [1] can be downloaded using the scipt in AASIST [2] repository
python ./aasist/download_dataset.py
Speaker & spoofing embedding extraction
Speaker embeddings and spoofing embeddings can be extracted using below script. Extracted embeddings will be saved in ./embeddings.
- Speaker embeddings are extracted using the ECAPA-TDNN [3].
- Implmented by https://github.com/TaoRuijie/ECAPATDNN
- Spoofing embeddings are extracted using the AASIST [2].
- We also prepared extracted embeddings.
- To use prepared emebddings, git-lfs is required. Please refer to https://git-lfs.github.com for further instruction. After installing git-lfs use following command to download the embeddings.
git-lfs install git-lfs pull
- To use prepared emebddings, git-lfs is required. Please refer to https://git-lfs.github.com for further instruction. After installing git-lfs use following command to download the embeddings.
python save_embeddings.py
Baseline 2 Training
Run below script to train Baseline2 in the evaluation plan.
- It will reproduce Baseline2 described in the Evaluation plan.
python main.py --config ./configs/baseline2.conf
Developing own models
- Currently adding...
Adding custom DNN architecture
- create new file under
./models/. - create a new configuration file under
./configs - in the new configuration, modify
model_archand add required arguments inmodel_config. - run
python main.py --config {USER_CONFIG_FILE}
Using only metrics
Use get_all_EERs in metrics.py to calculate all three EERs.
- prediction scores and keys should be passed on using
protocols/ASVspoof2019.LA.asv.dev.gi.trl.txtorprotocols/ASVspoof2019.LA.asv.eval.gi.trl.txt
References
[1] ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech
@article{wang2020asvspoof,
title={ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech},
author={Wang, Xin and Yamagishi, Junichi and Todisco, Massimiliano and Delgado, H{\'e}ctor and Nautsch, Andreas and Evans, Nicholas and Sahidullah, Md and Vestman, Ville and Kinnunen, Tomi and Lee, Kong Aik and others},
journal={Computer Speech \& Language},
volume={64},
pages={101114},
year={2020},
publisher={Elsevier}
}
[2] AASIST: Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks
@inproceedings{Jung2022AASIST,
author={Jung, Jee-weon and Heo, Hee-Soo and Tak, Hemlata and Shim, Hye-jin and Chung, Joon Son and Lee, Bong-Jin and Yu, Ha-Jin and Evans, Nicholas},
booktitle={Proc. ICASSP},
title={AASIST: Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks},
year={2022}
[3] ECAPA-TDNN: Emphasized Channel Attention, propagation and aggregation in TDNN based speaker verification
@inproceedings{desplanques2020ecapa,
title={{ECAPA-TDNN: Emphasized Channel Attention, propagation and aggregation in TDNN based speaker verification}},
author={Desplanques, Brecht and Thienpondt, Jenthe and Demuynck, Kris},
booktitle={Proc. Interspeech 2020},
pages={3830--3834},
year={2020}
}
56 Jan 2, 2023
1 Nov 20, 2021
22 Nov 10, 2022
3 Aug 20, 2022
48 Dec 30, 2022
47 Dec 19, 2022
47 Jan 1, 2023
9 Dec 21, 2022
208 Dec 24, 2022
11 Dec 26, 2022
79 Oct 08, 2022
1.8k Dec 30, 2022
12 Nov 09, 2022
[email protected]">
14 Sep 10, 2022
351 Nov 18, 2022
6.9k Jan 01, 2023
3 Jan 08, 2023
43 Oct 07, 2022
8 Oct 26, 2022
1 Feb 11, 2022
51 Jun 24, 2022
1.1k Jan 03, 2023
5 Apr 14, 2022
100 Dec 31, 2022
226 Dec 30, 2022
520 Dec 26, 2022
41 Sep 17, 2022
8 Dec 04, 2022
4.5k Dec 31, 2022