Naturalness-Aware Curriculum Learning with Dynamic Temperature for Speech Deepfake Detection

This repository contains our implementation and pretrained models of the paper accepted by Interspeech 2025.

• Title: Naturalness-Aware Curriculum Learning with Dynamic Temperature for Speech Deepfake Detection
• Author: Taewoo Kim, Guisik Kim, Choongsang Cho, Young Han Lee
• Affiliation: Korea Electronics Technology Institute (KETI)
• Paper

Environment

• Docker: nvidia/cuda:11.3.1-cudnn8-devel-ubuntu20.04
• Device: NVIDIA A40 48GB GPU

Installation

1. Install pyenv
2. Clone this repository
3. Setup virtual environment and install python requirements.

pyenv install 3.8.0
pyenv virtualenv 3.8.0 nacl_sdd
pyenv local nacl_sdd
poetry env use python
poetry install

4. The Wav2vec2.0 model (XLS-R 300M) can be downloaded from here.

Data Preparation

1. Downlaod datasets

   • ASVSpoof2019: Training and validation set
   • ASVSpoof2021: Evaluation set 1
   • InTheWild: Evaluation set 2

2. Download keys for evaluation

   • LA keys
   • PA keys
   • DF keys

3. Arrange datasets

├── datasets
│   ├── ASVspoof2019_LA_train
│   │   └── wav/
│   ├── ASVspoof2019_LA_dev
│   │   └── wav/
│   ├── ASVspoof2021_LA_eval
│   │   └── wav/
│   ├── ASVspoof2021_DF_eval
│   │   └── wav/
│   ├── in_the_wild
│   │   └── wav/
│   │
│   ├── keys
│   │   └── DF/
│   │   └── LA/
│   │   └── PA/
│   │   └── ITW/
│   ├── protocols
│   ├── utmos
│
├── pretrained
├── results

Pretrained Model and Results

Downlaod pretrained XLS-R Conformer models and results.

wget -O download.zip https://bit.ly/interspeech2025_NACL && unzip download.zip

Usage

# Training
sh scripts/train.sh

# Evaluation for LA & DF
sh scripts/evaluate.sh

# Evaluation for ITW
sh scripts/evaluate_itw.sh

Results

# Calculate Score for LA & DF
poetry run python common/eval/main.py --cm-score-file results/table_1/LA_fix/conformer_cl_dt.txt --track LA --subset eval
poetry run python common/eval/main.py --cm-score-file results/table_1/DF_fix/conformer_cl_dt.txt --track DF --subset eval

# Calculate Score for ITW (Table 3)
poetry run python common/eval/evaluate_in_the_wild.py results/table_3/ITW/conformer_cl_dt.txt datasets/keys/ITW/meta.csv

Citation

@inproceedings{kim2025naturalness,
  author={Kim, Taewoo and Kim, Guisik and Cho, Choongsang and Lee, Young Han},
  booktitle={arXiv preprint arXiv:2505.13976},
  title={Naturalness-Aware Curriculum Learning with Dynamic Temperature for Speech Deepfake Detection},
  year={2025}
}