Statistical features for detection of voice spoofing

Code for the paper "Bag of features for voice anti-spoofing"

We introduce a ”Bag of features”: a large number of different features for synthesized voice detection. "Bag of features" consist of a bunch of statistical parameters calculated on the raw audio signal and various spectrograms generated from it. We developed anti-spoofing system based on the introduced set of features that demonstrates outstanding results on ASVspoof 2019 challenge LA section as a single system giving a 3.93% equal error rate (EER) on the evaluation set.

Setup

Where are 2 options to clone the repository:

• Code only

GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/IDRnD/antispoofing-features.git

• Code + pretrained models and precomputed features

git clone https://github.com/IDRnD/antispoofing-features.git

Then install all required dependencies:

cd antispoofing-features
pip install -r requirements.txt

• Setup path for downloading and extracting of the ASVspoof 2019 dataset (dataset_path variable in the config.py)
• Run the next script for downloading and extraction of the dataset:

python download_dataset.py

• Setup number of processes used for parallel computations (8 by default)

Extraction of features

For extraction of statistical features run the next script:

python extract_features.py

The script outputs extracted features to the data directory:

data
|__dev
   |__repeats.npy
   |__stats.npy
    ...
    
|__train
   |__repeats.npy
   |__stats.npy
    ...
   
|__val
   |__repeats.npy
   |__stats.npy
    ...

Training the model

For training of decision tree-based models on the top of generated features run the next script:

python train_pipeline.py

Trained models will be saved to the models directory if save_models parameter of config is set to True.

Note

If you have a GPU with CUDA support you can use it for acceleration of training process. use_gpu parameter of config should be set to True (False by default, also check gpu_device_id parameter).

Evaluation

For evaluation of the EER score on the validation set of ASVspoof 19 LA dataset use model_testing.ipynb notebook.

Citation

If you find this code useful please cite us in your work:

@article{Torgashov2020BagOfFeatures,
  title={Bag of features for voice anti-spoofing},
  author={Nikita Torgashov, Ivan Iakovlev and Konstantin Simonchik},
  booktitle = {submitted to Interspeech},
  year={2020}
}