This repository includes the code to reproduce our paper Raw Differentiable Architecture Search for Speech Deepfake and Spoofing Detection published in the ASVspoof 2021 workshop.
pip install -r requirements.txt
Codes were tested with python==3.8.8 and CUDA Version==11.2
The ASVspoof 2019 database can be downloaded from here
The extracted data should be orginased as:
- LA/
- ASVspoof2019_LA_dev/flac/...
- ASVspoof2019_LA_eval/flac/...
- ASVspoof2019_LA_train/flac/...
- ASVspoof2019.LA.cm.dev.trl.txt
- ASVspoof2019.LA.cm.eval.trl.txt
- ASVspoof2019.LA.cm.train.trn.txt
- ASVspoof2019.LA.cm.train.trn_h.txt (uploaded in /split_protocols)
- ASVspoof2019.LA.cm.train.trn_t.txt (uploaded in /split_protocols)
- ...
For convience, you can change the codes' default --data argument to '/path/to/your/LA', instead of typing it for each run.
To search with 8 layers with 64 initial channels, and with fixed sinc layer initialised using mel scale:
python train_search.py --layers=8 --init_channels=64 --sinc_scale=mel
You can also try with other two scales (linear scale: linear and inverse-mel scale: lem), and set the scale to learnable:
python train_search.py --layers=8 --init_channels=64 --sinc_scale=linear --trainable
To train with the reported best architecture in the paper, using 8 layers, 64 initial channels and masked mel scale sinc layer:
python train_model.py --arch=ARCH --layers=8 --init_channels=64 --sinc_scale=mel
replace ARCH with "Genotype(normal=[('dil_conv_5', 1), ('dil_conv_3', 0), ('dil_conv_5', 1), ('dil_conv_5', 2), ('std_conv_5', 2), ('skip_connect', 3), ('std_conv_5', 2), ('skip_connect', 4)], normal_concat=range(2, 6), reduce=[('max_pool_3', 0), ('std_conv_3', 1), ('dil_conv_3', 0), ('dil_conv_3', 2), ('skip_connect', 0), ('dil_conv_5', 2), ('dil_conv_3', 0), ('avg_pool_3', 1)], reduce_concat=range(2, 6))"
To train with the reported best architecture in the paper, using 8 layers, 64 initial channels and masked, also learned linear scale sinc layer:
python train_model.py --arch=ARCH_trainable_linear --layers=8 --init_channels=64 --sinc_scale=linear --pre_trained=PATH
replace ARCH_trainable_linear with the corresponding architecture and replace PATH with the path of the pre-trained model saved during architecture search (for example, pre_trained_models/trainable_linear_in_search.pth). Please contact the first author if you need these files.
Because GRU layers are used in Raw-PC-DARTS, to keep the results exactly same (espically the initial architecture of searching stage) with a fixed seed, please try:
CUBLAS_WORKSPACE_CONFIG=:16:8 python train_search/model.py --seed=0
See the official document for this at here.
To evaluate the saved model using the same architecture in train from scratch on LA Evaluation partition:
python evaluate.py --arch=ARCH --model=PATH --layers=8 --init_channels=64 --frontend=SCALE
also replace ARCH with the corresponding architecture, PATH with /path/to/your/saved/models/epoch_x.pth, and SCALE with the corresponding scale.
The pre-trained models are too large to upload, so only the final scores are uploaded under /scores. You can contact the first author for the saved models through surname AT eurecom DOT fr.
The pre-trained models are uploaded to here.
If you find this repository useful, please consider citing:
@inproceedings{ge21_asvspoof,
author={Wanying Ge and Jose Patino and Massimiliano Todisco and Nicholas Evans},
title={{Raw Differentiable Architecture Search for Speech Deepfake and Spoofing Detection}},
year=2021,
booktitle={Proc. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge},
pages={22--28},
}
This work is supported by the ExTENSoR project funded by the French Agence Nationale de la Recherche (ANR).
Codes are based on the implementations of pc-darts-anti-spoofing , rawnet2-antispoofing and project-NN-Pytorch-scripts.