Quick Start

ASVspoof2019 LA, single GPU

  1. Prepare data

  • Download ASVspoof2019 LA and set ASVspoof_dir to its root.

  • From project root:

cd egs/detection/asvspoof2019/v03_resnet18
bash run.sh --stage 1 --stop_stage 2 --ASVspoof_dir /path/to/ASVspoof2019_LA

It will use egs/detection/asvspoof2019/v03_resnet18/run.sh stages 1–2 to build lists:

  • data/asvspoof2019/{train,dev,eval}/wav.scp

  • data/asvspoof2019/{train,dev,eval}/utt2lab

  • data/asvspoof2019/{train,dev,eval}/raw.list or shard.list

  • local/prepare_data.sh populates wav.scp/utt2lab; tools/make_raw_list.py or tools/make_shard_list.py creates the lists consumed by training.

  • For other datasets, prepare the same files under data/<dataset>/<split>/ following the ASVspoof2019 layout.

  1. Train a baseline

bash run.sh --stage 3 --stop_stage 3 --gpus "[0]" --ASVspoof_dir /path/to/ASVspoof2019_LA

Expected log snippet (from exp_dir/train.log):

INFO exp_dir is: exp/ResNet18-i5p5-smallWeightDecay-earlystop
INFO Train epoch iteration number: 820
INFO Epoch 1/50 step 100 loss=0.48 acc=0.87 lr=1.0e-3

  1. Evaluate (embeddings -> logits -> llr -> metrics)

bash run.sh --stage 4 --stop_stage 7 --gpus "[0]" --ASVspoof_dir /path/to/ASVspoof2019_LA