CodecFake Source Tracing

Official code for our source tracing work on codec-based deepfake speech. Pretrained model checkpoints are hosted on Hugging Face.

🏆 Towards Generalized Source Tracing for Codec-Based Deepfake Speech received a Best Student Paper Nomination (Top 2.5%) at IEEE ASRU 2025.

🛠️ Setup

Dataset Download

Download the CodecFake+ dataset from Hugging Face.

CodecFake+/
├── all_data_16k/          # CoRS + maskgct_vctk set
│   ├── p225_001_audiodec_24k_320d.wav
│   ├── p225_001_bigcodec.wav
│   ├── ....
│   └── s5_400_xocdec_hubert_general_audio.wav
└── SLMdemos_16k/          # CoSG set
    ├── SIMPLESPEECH1/     
    ├── VIOLA/
    ├── ....
    └── MASKGCT/

Note: The tree above and the metadata/ files in this repository use the file names from our internal copy of the dataset. The public release on Hugging Face uses a different layout (Codecfake_plus_CoRS/ and CoSG/) and different CoSG file names, so the paths need to be adapted before running the scripts.

Pretrained Weights Setup

• For Wav2Vec2-AASIST

  • Place xlsr2_300m.pt directly into w2v2_aasist_baseline/

• For SAST Net

  • Create directory Pretrain_weight inside SAST_Net/
  • Download and place the following checkpoints in SAST_Net/Pretrain_weight:

  Model	Description	Download
  xlsr2_300m.pt	Wav2Vec2 pretrained weight	Download
  mae_pretrained_base.pth	AudioMAE pretrained on AudioSet	Download
  tuned_weight.pth	Wav2Vec2-AASIST on CodecFake+	Download

Environment Setup

conda env create -f environment.yml
conda activate CodecFakeSourceTracing

---

🚀 Inference

Notation

• Tasks

  • BIN: Binary spoof detection task
  • VQ: Vector quantization source tracing task
  • AUX: Auxiliary training objective source tracing task
  • DEC: Decoder type source tracing task

• Training Subsets

  • vq: VQ taxonomy sampling (MVQ : SVQ : SQ = 1:1:1)
  • aux: AUX taxonomy sampling (None : Semantic Distillation : Disentanglement = 1:1:1)
  • dec: DEC taxonomy sampling (Time : Freqency = 1:1)

Model Checkpoints

• Wav2Vec2-AASIST

  Single-Task Learning Models

  Model	Task	Trained Dataset	Download Links
  S_BIN	BIN	vq / aux / dec	vq • aux • dec
  S_VQ	VQ	vq	Download
  S_AUX	AUX	aux	Download
  S_DEC	DEC	dec	Download

  Dual-Task Learning Models

  Model	Task	Trained Dataset	Download Links
  D_VQ	BIN / VQ	vq	Download
  D_AUX	BIN / AUX	aux	Download
  D_DEC	BIN / DEC	dec	Download

  Multi-Task Learning Models

  Model	Task	Trained Dataset	Download Links
  M1	BIN / VQ / AUX / DEC	vq / aux / dec	vq • aux • dec
  M2	VQ / AUX / DEC	vq / aux / dec	vq • aux • dec

• SAST Net

  Model	Task	Trained Dataset	Download Links
  SAST Net	BIN	vq / aux / dec	vq • aux • dec
  	VQ	vq	Download
  	AUX	aux	Download
  	DEC	dec	Download

Running Inference

• Wav2Vec2-AASIST

  cd w2v2_aasist_baseline/
  bash inference.sh ${dataset_type} ${base_dir} ${checkpoint_path} ${model_type}

  Parameters:

  • dataset_type: "CoRS" or "CoSG"
  • base_dir: Path to dataset directory
    • For CoRS: "CodecFake+/all_data_16k/"
    • For CoSG: "CodecFake+/SLMdemos_16k/"
  • checkpoint_path: Path to model checkpoint
  • model_type: S_BIN / S_VQ / S_AUX / S_DEC / D_VQ / D_AUX / D_DEC / M1 / M2

• SAST Net

  cd SAST_Net/
  bash inference.sh ${base_dir} ${dataset_type} ${checkpoint_path} ${task} ${eval_output}

  Parameters:

  • base_dir: Path to dataset directory
  • dataset_type: "CoRS" or "CoSG"
  • checkpoint_path: Path to model checkpoint
  • task: Bin / AUX / DEC / VQ
  • eval_output: Results directory (default: "./Result")

---

🎯 Training

• Wav2Vec2-AASIST

  cd w2v2_aasist_baseline/
  bash train.sh ${base_dir} ${batch_size} ${num_epochs} ${lr} ${model_type} ${sampling_strategy}

  Parameters:

  • base_dir: Path to "CodecFake+/all_data_16k/"
  • batch_size: Batch size (default: 8)
  • num_epochs: Training epochs (default: 20)
  • lr: Learning rate (default: 1e-06)
  • model_type: S_BIN / S_VQ / S_AUX / S_DEC / D_VQ / D_AUX / D_DEC / M1 / M2
  • sampling_strategy: VQ / AUX / DEC

• SAST Net

  cd SAST_Net
  bash train.sh ${base_dir} ${save_dir} ${batch_size} ${num_epochs} ${lr} ${task} ${sampling_strategy} ${mask_ratio}

  Parameters:

  • base_dir: Path to "CodecFake+/all_data_16k/"
  • save_dir: Checkpoint save directory (default: ./models_SAST_Net)
  • batch_size: Batch size (default: 12)
  • num_epochs: Training epochs (default: 40)
  • lr: Learning rate (default: 5e-06)
  • task: Bin / VQ / AUX / DEC
  • sampling_strategy: VQ / AUX / DEC
  • mask_ratio: MAE mask ratio (default: 0.4)

📚 Citation

If this work helps your research, please consider citing our papers:

@inproceedings{chen2025codec,
  title     = {Codec-Based Deepfake Source Tracing via Neural Audio Codec Taxonomy},
  author    = {Chen, Xuanjun and Lin, I-Ming and Zhang, Lin and Du, Jiawei and Wu, Haibin and Lee, Hung-yi and Jang, Jyh-Shing Roger},
  booktitle = {{INTERSPEECH 2025}},
  pages     = {1538--1542},
  year      = {2025},
  doi       = {10.21437/Interspeech.2025-1297}
}

@inproceedings{chen2025towards,
  title     = {Towards Generalized Source Tracing for Codec-Based Deepfake Speech},
  author    = {Chen, Xuanjun and Lin, I-Ming and Zhang, Lin and Wu, Haibin and Lee, Hung-yi and Jang, Jyh-Shing Roger},
  booktitle = {2025 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
  pages     = {1--8},
  year      = {2025},
  doi       = {10.1109/ASRU65441.2025.11434771}
}

@article{chen2026codecfake+,
  title   = {{CodecFake+}: Codec-Based Resynthesized Data as a Proxy for Detecting {CodecFake} Speech},
  author  = {Chen, Xuanjun and Du, Jiawei and Wu, Haibin and Zhang, Lin and Lin, I-Ming and Chiu, I-Hsiang and Ren, Wenze and Tseng, Yuan and Tsao, Yu and Jang, Jyh-Shing Roger and Lee, Hung-yi},
  journal = {IEEE Transactions on Audio, Speech and Language Processing},
  volume  = {34},
  pages   = {2929--2944},
  year    = {2026},
  doi     = {10.1109/TASLPRO.2026.3692291}
}