Multi-Task-Learning-Improves-Synthetic-Speech-Detection

This code is for our accepted manuscript to 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP).

Data Preprocessing

Downloading the ASVspoof 2019 Logic Access Dataset.
Matlab is required.
Edit ./extract_feature/process_LA_data.m according to the absolute path of the dataset.

cd extract_feature
CUDA_VISIBLE_DEVICES=0 /data/users/yangli/Matlab/bin/matlab -nodesktop -nosplash -r process_LA_data.m
cd ..
python reload_data.py

Training

Install required Python packages:

pip install -r requirement.txt

Use our multi-task learning methods to train a synthetic speech detection network (Taking oc-softmax loss function as an example.).
vanilla

CUDA_VISIBLE_DEVICES=0,1 python3 train.py --add_loss ocsoftmax -o  ./models/ocsoftmax_vanilla -f /data/users/yangli/AIR-ASVspoof-master/LAfeatures/

+bonafide speech reconstruction

CUDA_VISIBLE_DEVICES=0,1 python3 train.py --add_loss ocsoftmax -o  ./models/ocsoftmax_recon_04 -f /data/users/yangli/AIR-ASVspoof-master/LAfeatures/ --S1 --lambda_r 0.04

+spoofing voice conversion

CUDA_VISIBLE_DEVICES=0,1 python3 train.py --add_loss ocsoftmax -o  ./models/ocsoftmax_conver_0003 -f /data/users/yangli/AIR-ASVspoof-master/LAfeatures/ --S2 --lambda_c 0.0003

+speaker classification

CUDA_VISIBLE_DEVICES=0,1 python3 train.py --add_loss ocsoftmax -o  ./models/ocsoftmax_class_00005 -f /data/users/yangli/AIR-ASVspoof-master/LAfeatures/ --S3 --lambda_m 0.00005

Combining all auxillary subtasks

CUDA_VISIBLE_DEVICES=0,1 python3 train.py --add_loss ocsoftmax -o  ./models/ocsoftmax_class_00005_recon_04_conver_0003 -f /data/users/yangli/AIR-ASVspoof-master/LAfeatures/ --S3 --lambda_m 0.00005 --S1 --lambda_r 0.04 --S2 --lambda_c 0.0003

Evaluation

An example:

CUDA_VISIBLE_DEVICES=0,1 python3 test.py -m ./models/ocsoftmax_class_00005_recon_04_conver_0003  -l ocsoftmax --gpu 0 -f /data/users/yangli/AIR-ASVspoof-master/LAfeatures/

For details, please refer to test.py.

Checkpoint Downloading

Google Drive

Cite this

@inproceedings{mo2022multi,
  title={Multi-Task Learning Improves Synthetic Speech Detection},
  author={Mo, Yichuan and Wang, Shilin},
  booktitle={ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={6392--6396},
  year={2022},
  organization={IEEE}
}