AIR-ASVspoof

This repository contains the official implementation of "One-class Learning Towards Synthetic Voice Spoofing Detection". [paper] [arXiv] [video]
You can find a brief explanation here.

Requirements

Install python 3.9 and the latest version of pytorch (with/without CUDA, depending on your device).
If you encounter any problems with python, try installing python==3.6 and pytorch==1.1.0. (Required for Training of data)
Note:- Pytorch is currently compatible with only python versions from 3.7 to 3.9. So, that might cause some errors in torch packages used for train.py and others.
But, you can run the testing part with the already provided pre-trained model.
Check this issue for more information: yzyouzhang#36 (The inputs of test should be anti-spoofing_lfcc_model.pt and anti-spoofing_loss_model.pt, output should be checkpoint_cm_score.txt)

Downloads

If you want to work on this project, you will have to download a few things.

1. An IDE(IntelliJ) to run .py files and MATLAB to run .m files (feature_extraction.py is the python version of process_LA_data.m file.).
2. After installing python, download CUDA if you have supporting NVIDIA gpu (for faster testing and training), and download pytorch with/without CUDA accordingly from here.
3. For Data Preparation, database of audio samples (LA.zip) should be downloaded from here.
   Extract LA.zip into DS_10283_3336(Dataset) folder.
   MATLAB implementations for LFCC features are available here.

Data Preparation

The LFCC features are extracted with the MATLAB implementation provided by the ASVspoof 2019 organizers.
First run the process_LA_data.m with MATLAB, and then run python reload_data.py with python.
If you don't have/want to use MATLAB, use the python file feature_extraction.py, instead of process_LA_data.m, then run reload_data.py.
Make sure you change the directory path to the path on your machine.

For successful execution of process_LA_data.m, we will have to create a folder to contain the output .m files.
Create a folder anti-spoofing inside DS_10283_3336 as shown in pathToFeatures line of the process_LA_data.m file.
The output folder there in process_LA_data.m is the Features folder. It should be further divided into dev, eval and train folders.

Run the training code

Before running the train.py, please change the path_to_database, path_to_features, path_to_protocol according to the files location on your machine.
path_to_database is the path to your DS_10283_3336 folder.
path_to_features is the path to the output folder created for Data Preperation part above.
path_to_protocol is the path to ASVspoof2019_LA_cm_protocols folder.

python train.py --add_loss ocsoftmax -o D:/Programming/Python/Python/AIR-ASVspoof-Suchit/models1028/ocsoftmax/test_results --gpu 0

Change --gpu 0 to --gpu 1, if you want to run the code with your nvidia gpu.

Run the test code with trained model

You can change the model_dir to the location of the trained model you would like to test with.
Make sure that you change all the required paths in the code to your folder paths.

python test.py -m D:/Programming/Python/Python/AIR-ASVspoof-Suchit/models1028/ocsoftmax/test_results -l ocsoftmax

The output will be inside the text file checkpoint_cm_score.txt in model_dir folder.
The output will be in the form of: LA_E_2834763 A11 spoof -0.9655166864395142.

The final output of the model with a min t-DCF of 0.059, and EER of 2.19%. This model could be in the 3rd position of ASVspoof 2019 competition's LA subset.

You can find citations and future works in the author's repository here.