SONAR: A Synthetic AI-Audio Detection Framework and Benchmark

An official implementation of "SONAR: A Synthetic AI-Audio Detection Framework and Benchmark"

Enviroment

• conda env export > environment.yml
• conda activate sonar

Datasets

Please download the following datasets and extract them in /data/ or change the database path correspondingly in the code.

• Wavefake
• LibriSeVoc
• In-the-wild
• LJ-Speech
• SONAR

The structure should look like

data
├── LJSpeech-1.1
│ 	├── wavs
│	├── metadata.csv
│ 	└── README
├── wavefake
│ 	├── ljspeech_full_band_melgan
│	├── ljspeech_hifiGAN
│	├── ...
│ 	└── ljspeech_waveglow
├── LibriSeVoc
│ 	├── diffwave
│	├── gt
│	├── ...
│ 	└── wavernn
├── in_the_wild
│ 	├── 0.wav
│	├── ...
│	├── 31778.wav
│ 	└── meta.csv

Usage examples

To train traditional models, please run main_tm.py

Arguments

• --config: config files for different models.

• Train AASIST on wavefake

  python main_tm.py --config ./config/AASIST.conf 
  

• Evaluation (modify the model_path in corresponding config files.)

  python main_tm.py --config ./config/AASIST.conf --eval
  

To fine-tune foundation models, please run main_fm.py

• Fine-tune Wave2Vec2BERT

  python main_fm.py --model wave2vec2bert
  

Acknowledgements

This repository is built on top of the following open source projects.

• AASIST
• wavefake