Speaker-Anti-Spoofing-Classifiers

This repository provides a basic speaker anti-spoofing system using neural networks in pytorch 1.0+ and python 3.6+.

Python requirements are:

pandas==0.25.3
tqdm==4.28.1
torch==1.3.1
matplotlib==3.1.1
numpy==1.17.4
fire==0.2.1
pytorch_ignite==0.2.1
h5py==2.10.0
six==1.13.0
adabound==0.0.5
ignite==1.1.0
librosa==0.7.1
metrics==0.3.3
pypeln==0.1.10
PyYAML==5.2
scikit_learn==0.22

Moreover, to download the data, you will need wget.

Evaluation scripts are directly taken from the baseline of the ASVspoof2019 challenge, seen here

Datasets

The most broadly used datasets for spoofing detection (currently) are:

• ASVspoof2019 encompassing logical and physical attacks alike.
• ASVspoof2017 encompassing only physical attacks with the focus on in the wild devices and scenes.
• ASVspoof2015 encompassing only logical attacks with the focus on synthesize and voice conversion attacks.

For mixed logical and physical attacks, the mixed AVspoof dataset ( a subpart of it is the BTAS16 dataset) can be also used. The dataset is publicly available, but only for research purposes.

Moreover, rather recent the Fake-or-Real (FoR) dataset was introduced using openly available synthesizers such as Baidu, Google, Amazon to create spoofs for logical access.

Feature extraction

Features are extracted using the librosa toolkit. We provide four commonly used features: CQT,(Linear)log-Spectrograms, log-Mel-Spectrograms and raw wave features.

Models

Currently, the (what I think) most popular model is LightCNN, which is the winner of the ASVspoof2017 challenge paper. Another model called CGCNN, which modified the MFM activation to gated linear unit (GLU) activations has been successfully employed in the ASVspoof2019 challenge. My current implementation can be seen in models.py.

All experiments conducted in this repository use 90% of the training set as training data and 10% as cross-validation. Development data is not used at all. All results are only evaluated on the respective evaluation set.

The baseline results are as follows:

Dataset	Feature	Model	EER
ASV15	Spec	LightCNN	0.68
ASV15	Spec	CGCNN	0.33
ASV17	Spec	LightCNN	12.18
ASV17	Spec	CGCNN	11.16
BTAS16	Spec	LightCNN	2.04
BTAS16	Spec	CGCNN	3.26
FoR-norm	Spec	LightCNN	5.43
FoR-norm	Spec	CGCNN	18.64

Config

The framework here uses a combination of google-fire and yaml parsing to enable a convenient interface for anti-spoofing model training. By default one needs to pass a *.yaml configuration file into any of the command scripts. However parameters of the yaml files can also be overwritten on the fly:

python3 run.py train config/asv17.yaml --model MYMODEL, searches for the model MYMODEL in models.py and runs the experiment using that model.

Other notable arguments are:

• --model_args '{filters:[60,60]}' sets the filter sizes of a convolutional model to 60, 60.
• --batch_size 32 --num_workers 2' sets training hyper parameters batch_size as well as the number of async workers for dataloading.
• --transforms '[timemask,freqmask]' applies augmentation on the training data, defined in augment.py.

Commands

The main script of this repository is run.py. Five commands are available:

• train e.g., python3 run.py train config/asv17.yaml trains a specified model on specified data.
• score e.g., python3 run.py score EXPERIMENT_PATH OUTPUTFILE.tsv --testlabel data/filelists/asv17/eval.tsv --testdata data/hdf5/asv17/spec/eval.h5 scores a given experiment and produces OUTPUTFILE.tsv containing the respective scores. End-to-End scoring is utilized, where the genuine class scores are representative of the model belief. Only a single dataset can be scored.
• evaluate_eer uses the library contained in evaluation/ to calculate an EER. Example usage is: python3 run.py evaluate_eer experiments/asv17/LightCNN/SOMEEXPERIEMNT/scored.txt data/filelists/asv17/eval.tsv output.txt. output.txt is generated with the results ( which are also printed to console ).
• run e.g., python3 run.py run config/asv17.yaml trains, scores and evaluates an experiment. Can also support multiple tests using --testlabel ['a.tsv','b.tsv]' or just updating the config file. Is effectively train, score, evaluate_eer in one and the most recommended way of running any experiment.

Usage

For a simple e.g., ASVspoof2017 dataset run, please run the following:

git clone https://github.com/RicherMans/Speaker-Anti-Spoofing-Classifiers
cd Speaker-Anti-Spoofing-Classifiers
pip3 install -r requirements.txt
cd data/scripts
bash download_asv17.sh
bash prepare_asv17.sh
cd ../../features/
python3 extract_feature.py ../data/filelists/asv17/train.tsv -o hdf5/asv17/spec/train.h5 # Extracts spectrogram features
python3 extract_feature.py ../data/filelists/asv17/eval.tsv -o hdf5/asv17/spec/eval.h5 #Spectrogram features
cd ../
python3 run.py run config/asv17.yaml # Runs LightCNN Model. Results will be displayed in the console and a directory experiments/asv17 will be created.
python3 run.py run config/asv17.yaml --model CGCNN # Runs CGCNN Model