Skip to content

Latest commit

 

History

History
executable file
·
96 lines (86 loc) · 3.72 KB

README.md

File metadata and controls

executable file
·
96 lines (86 loc) · 3.72 KB
Raw

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].
  • 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

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

  1. create new file under ./models/.
  2. create a new configuration file under ./configs
  3. in the new configuration, modify model_arch and add required arguments in model_config.
  4. 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.txt or
    • protocols/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}
}