lqbbb / wav2vec2.0_DID

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation

This repository contains our implementation of the paper published in the Speaker Odyssey 2022 workshop, "Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation".

Paper link here

Installation

First, clone the repository locally, create and activate a conda environment, and install the requirements :

$ git clone https://github.com/TakHemlata/SSL_Anti-spoofing.git
$ conda create -n SSL_Spoofing python=3.7
$ pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
$ cd fairseq-a54021305d6b3c4c5959ac9395135f63202db8f1
(This fairseq folder can also be downloaded from https://github.com/pytorch/fairseq/tree/a54021305d6b3c4c5959ac9395135f63202db8f1)
$ pip install --editable ./
$ pip install -r requirements.txt

Experiments

Dataset

Our experiments are performed on the logical access (LA) and deepfake (DF) partition of the ASVspoof 2021 dataset (train on 2019 LA training and evaluate on 2021 LA and DF evaluation database).

The ASVspoof 2019 dataset, which can can be downloaded from here.

The ASVspoof 2021 database is released on the zenodo site.

LA here

DF here

For ASVspoof 2021 dataset keys (labels) and metadata are available here

Pre-trained wav2vec 2.0 XLSR (300M)

Download the XLSR models from here

Training LA

To train the model run:

CUDA_VISIBLE_DEVICES=0 main_SSL_LA.py --track=LA --lr=0.000001 --batch_size=14 --loss=WCE  

Testing LA and DF

To evaluate your own model on LA and DF evaluation dataset:

CUDA_VISIBLE_DEVICES=0 python main_SSL_LA.py --track=LA --is_eval --eval --model_path='/path/to/your/best_SSL_model_LA.pth' --eval_output='eval_CM_scores_file_SSL_LA.txt'

CUDA_VISIBLE_DEVICES=0 python main_SSL_DF.py --track=DF --is_eval --eval --model_path='/path/to/your/best_SSL_model_LA.pth' --eval_output='eval_CM_scores_file_SSL_DF.txt'

We also provide a pre-trained models. To use it you can run:

Pre-trained SSL antispoofing models are available here

CUDA_VISIBLE_DEVICES=0 python main_SSL_LA.py --track=LA --is_eval --eval --model_path='/path/to/Pre_trained_models/best_SSL_model_LA.pth' --eval_output='eval_pre_trained_model_CM_scores_file_SSL_LA.txt'

Compute the min t-DCF and EER(%) on 2021 LA and DF evaluation dataset

python evaluate_2021_LA.py Score_LA.txt ./keys eval

python evaluate_2021_DF.py Score_DF.txt ./keys eval

Contact

For any query regarding this repository, please contact:

  • Hemlata Tak: tak[at]eurecom[dot]fr

Citation

If you use this code in your research please use the following citation:

@inproceedings{tak2022automatic,
  title={Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation},
  author={Tak, Hemlata and Todisco, Massimiliano and Wang, Xin and Jung, Jee-weon and Yamagishi, Junichi and Evans, Nicholas},
  booktitle={The Speaker and Language Recognition Workshop},
  year={2022}
}

About

License:MIT License


Languages

Language:Python 97.9%Language:Cuda 1.1%Language:C++ 0.6%Language:Cython 0.4%Language:Shell 0.0%