ash56/few_shot_spoof_detection

Few-Shot Synthetic Speech Detection

This repository contains the checkpoints and data manifests released for the GitHub project:

Ashigarg123/fewshot-spoof-detection

The models accompany Rapidly Adapting to New Voice Spoofing: Few-Shot Detection of Synthesized Speech Under Distribution Shifts and include few-shot prototypical networks, zero-shot SSL-AASIST classifiers, and the one-class Mahalanobis baseline.

Installation

Clone the code repository and install its dependencies:

git clone https://github.com/Ashigarg123/fewshot-spoof-detection.git
cd fewshot-spoof-detection
conda create -n fewshot-spoof python=3.10 -y
conda activate fewshot-spoof
pip install -r requirements.txt
pip install huggingface_hub

Follow the GitHub README to download the required XLS-R 300M frontend checkpoint and the evaluation audio.

Download Checkpoints

Download one checkpoint with huggingface_hub:

from huggingface_hub import hf_hub_download

checkpoint_path = hf_hub_download(
    repo_id="ash56/few_shot_spoof_detection",
    filename="checkpoints/binary_attn_5shot/best_model.pth",
)
print(checkpoint_path)

Or download all checkpoints, configs, and manifests:

from huggingface_hub import snapshot_download

release_dir = snapshot_download(
    "ash56/few_shot_spoof_detection",
    local_dir="checkpoints/released",
)
print(release_dir)

Released Models

Directory	GitHub evaluator	Model argument
checkpoints/binary_attn_5shot	src/evaluate.py	binary_attention
checkpoints/binary_mean_5shot	src/evaluate.py	binary_mean
checkpoints/multiclass_attn_5shot	src/evaluate.py	multiclass_attention
checkpoints/zero_shot_ssl_aasist_binary	src/zero_shot.py evaluate	binary
checkpoints/zero_shot_ssl_aasist_multiclass	src/zero_shot.py evaluate	multiclass
checkpoints/one_class_mahalanobis	src/mahalanobis.py evaluate	not applicable

For example, evaluate the binary self-attentive few-shot model after downloading the full snapshot to checkpoints/released:

python src/evaluate.py --model binary_attention \
  --ssl-checkpoint checkpoints/xlsr/xlsr2_300m.pt \
  --checkpoint checkpoints/released/checkpoints/binary_attn_5shot/best_model.pth \
  --eval-list checkpoints/released/manifests/evaluation/shiftyspeech/ja/apnet2.txt \
  --audio-root /path/to/ShiftySpeech \
  --output-dir runs/jsut_apnet2 \
  --support-sizes 5 10 \
  --num-iters 100 \
  --seed 7 \
  --device cuda:0

See the GitHub README for dataset setup and commands for every model family. XLS-R 300M and evaluation audio are not redistributed in this checkpoint repository.

Citation

If you use these checkpoints or the associated code, please cite:

@inproceedings{garg2025rapidly,
  title={Rapidly Adapting to New Voice Spoofing: Few-Shot Detection of Synthesized Speech Under Distribution Shifts},
  author={Garg, Ashi and Cai, Zexin and Xinyuan, Henry Li and Garc{\'\i}a-Perera, Leibny Paola and Khudanpur, Kevin Duh Sanjeev and Wiesner, Matthew and Andrews, Nicholas},
  booktitle={2025 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
  pages={1--7},
  year={2025},
  organization={IEEE}
}