Adapted version -- Baseline of the WildSpoof challenge SASV track.

• Adapted by Hye-jin Shim.

This version is an adaptation to be used as a baseline in the WildSpoof, Spoofing-robust Automatic Speaker Verification (SASV) track.

This baseline is based on the system proposed in a paper Towards single integrated spoofing-aware speaker verification embeddings, which was presented at Interspeech 2023. This adapted version trains a Deep Neural Network (DNN) that has two output layers, one for speaker identification (multi-class classification) and the other for anti-spoofing (bonafide/spoof classification).

Please send an email to shimhz6.6@gmail.com for questions related to this adapted version for WildSpoof challenge SASV track.

Prerequisites

Activate a conda environment

• Not mandatory, but we recommend to initialize a conda envionment and match the development environment. This is because Deep Neural Network (DNN) models' results are not deterministic when trained multiple times. Here is the environment we used.

conda create --name {wildspoof_sasv}
conda install -y conda "python=${3.9.19}"
conda install pytorch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 pytorch-cuda=12.1 -c pytorch -c nvidia\n

• Afterward, install the packages via pip install -r requirements.txt.

Models

This adaptation employs SKA-TDNN. You can also select other models implemented in the main branch (or your own model) using the --model option:

Training

Train using the command below.

CUDA_VISIBLE_DEVICES=0 python trainSASVNet.py \
  --max_frames 400 \
  --num_spk 400 \
  --num_utt 2 \
  --batch_size 40 \
  --trainfunc sasv_e2e_v1 \
  --optimizer adamW \
  --scheduler cosine_annealing_warmup_restarts \
  --lr_t0 8 \
  --lr_tmul 1.0 \
  --lr_max 1e-4 \
  --lr_min 0 \
  --lr_wstep 0 \
  --lr_gamma 0.8 \
  --margin 0.2 \
  --scale 30 \
  --num_class 1160 \
  --save_path exp/sasv_baseline \
  --train_list corpus/spoofceleb/metadata/train.csv \
  --eval_list corpus/spoofceleb/protocol/sasv_development_evaluation_protocol.csv \
  --train_path corpus/spoofceleb/flac/train \
  --eval_path corpus/spoofceleb/flac/development \
  --spk_meta_train spk_meta/spk_meta_trn_spoofceleb.pk \
  --spk_meta_eval spk_meta/spk_meta_dev_spoofceleb.pk \
  --musan_path /path/to/dataset/MUSAN/musan_split \
  --rir_path /path/to/dataset/RIRS_NOISES/simulated_rirs \
  --model SKA_TDNN

Evaluation

You can evaluate your checkpoint of a model using:

CUDA_VISIBLE_DEVICES=0 python trainSASVNet.py \
        --eval \
        --eval_frames 0 \
        --num_eval 1 \
        --eval_list corpus/spoofceleb/protocol/sasv_evaluation_evaluation_protocol.csv \
        --eval_path corpus/spoofceleb/flac/evaluation/ \
        --model SKA_TDNN \
        --initial_model /path/to/your_model/pretrained_weight.model

Metric

We use the Agnostic Detection Cost Function (a-DCF) as the main metric, the primary metric used in the challenge to rank the submissions.

Reference

@article{jung2025spoofceleb,
  title={SpoofCeleb: Speech deepfake detection and SASV in the wild},
  author={Jung, Jee-weon and Wu, Yihan and Wang, Xin and Kim, Ji-Hoon and Maiti, Soumi and Matsunaga, Yuta and Shim, Hye-jin and Tian, Jinchuan and Evans, Nicholas and Chung, Joon Son and others},
  journal={IEEE Open Journal of Signal Processing},
  year={2025},
  publisher={IEEE}
}

@article{jung2024text,
  title={The Text-to-speech in the Wild (TITW) Database},
  author={Jung, Jee-weon and Zhang, Wangyou and Maiti, Soumi and Wu, Yihan and Wang, Xin and Kim, Ji-Hoon and Matsunaga, Yuta and Um, Seyun and Tian, Jinchuan and Shim, Hye-jin and others},
  journal={ISCA Interspeech 2025},
  year={2024}
}

@inproceedings{mun2022frequency,
  title={Frequency and Multi-Scale Selective Kernel Attention for Speaker Verification},
  author={Mun, Sung Hwan and Jung, Jee-weon and Han, Min Hyun and Kim, Nam Soo},
  booktitle={Proc. IEEE SLT},
  year={2022}
}

@inproceedings{mun2023towards
  title={Towards single integrated spoofing-aware speaker verification embeddings},
  author={Mun, Sung Hwan and Shim, Hye-jin and Tak, Hemlata and Wang, Xin and Liu, Xuechen and Sahidullah, Md and Jeong, Myeonghun and Han, Min Hyun and Todisco, Massimiliano and Lee, Kong Aik and others},
  booktitle={Proc. Interspeech},
  year={2023}
}

@inproceedings{shim2024an
  title={a-DCF: an architecture agnostic metric with application to spoofing-robust speaker verification},
  author={Shim, Hye-jin and Jung, Jee-weon and Kinnunen, Tomi and Evans, Nicholas and Bonastre, Jean-Francois and Lapidot, Itshak},
  booktitle={Proc. Speaker Odyssey},
  year={2024}
}