Overview

This repository is an implementation of the Auxiliary Raw Net (ARawNet), which is ASVSpoof detection system taking both raw waveform and handcrafted features as inputs,to balance the trade-off between performance and model complexity. The paper can be checked here.

The model performance is tested on the ASVSpoof 2019 Dataset.

Setup

Environment

• Create a conda environment with conda env create -f environment.yml.
• Activate the conda environment with conda activate .

``

Data preprocessing

.
├── data                       
│   │
│   ├── PA                  
│   │   └── ...
│   └── LA           
│       ├── ASVspoof2019_LA_asv_protocols
│       ├── ASVspoof2019_LA_asv_scores
│       ├── ASVspoof2019_LA_cm_protocols
│       ├── ASVspoof2019_LA_train
│       ├── ASVspoof2019_LA_dev
│       
│
└── ARawNet

1. Download dataset. Our experiment is trained on the Logical access (LA) scenario of the ASVspoof 2019 dataset. Dataset can be downloaded here.

2. Unzip and save the data to a folder data in the same directory as ARawNet as shown in below.

3. Run python preprocess.py Or you can use our processed data directly under "/processed_data".

Train

python train_raw_net.py yaml/RawSNet.yaml --data_parallel_backend -data_parallel_count=2

Evaluate

python eval.py

Check Model Size and multiply-and-accumulates (MACs)

python check_model_size.py yaml/RawSNet.yaml

Model Performance

Accuracy metric

min t−DCF =min{βPcm (s)+Pcm(s)}

Explanations can be found here: t-DCF

Experiment Results

Cite Our Paper

If you use this repository, please consider citing:

@inproceedings{Teng2021ComplementingHF, title={Complementing Handcrafted Features with Raw Waveform Using a Light-weight Auxiliary Model}, author={Zhongwei Teng and Quchen Fu and Jules White and M. Powell and Douglas C. Schmidt}, year={2021} }

@inproceedings{Fu2021FastAudioAL, title={FastAudio: A Learnable Audio Front-End for Spoof Speech Detection}, author={Quchen Fu and Zhongwei Teng and Jules White and M. Powell and Douglas C. Schmidt}, year={2021} }