README.md

metadata

language:
  - en
pipeline_tag: audio-classification
datasets:
  - CodecFake/CodecFake_Plus_Dataset
tags:
  - audio
  - speech
  - deepfake-detection
  - audio-deepfake-detection
  - anti-spoofing
  - source-tracing
  - neural-audio-codec
  - codecfake

CodecFake Source Tracing Checkpoints

Pretrained model checkpoints for our source tracing work on codec-based deepfake speech. Training and inference code is on GitHub.

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

Tasks and Training Subsets

Task	Meaning
BIN	Binary spoof detection
VQ	Vector quantization source tracing
AUX	Auxiliary objective source tracing
DEC	Decoder type source tracing

Subset	Sampling in the training data
vq	MVQ : SVQ : SQ = 1 : 1 : 1
aux	None : Semantic Distillation : Disentanglement = 1 : 1 : 1
dec	Time : Frequency = 1 : 1

Checkpoints

Wav2Vec2-AASIST

Files are in model_checkpoints/Wav2Vec2_baseline/.

Model	Learning	Task	Trained on	Files
S_BIN	Single-task	BIN	vq / aux / dec	S_BIN_VQ_bal.pth, S_BIN_AUX_bal.pth, S_BIN_DEC_bal.pth
S_VQ	Single-task	VQ	vq	S_VQ.pth
S_AUX	Single-task	AUX	aux	S_AUX.pth
S_DEC	Single-task	DEC	dec	S_DEC.pth
D_VQ	Dual-task	BIN / VQ	vq	D_VQ.pth
D_AUX	Dual-task	BIN / AUX	aux	D_AUX.pth
D_DEC	Dual-task	BIN / DEC	dec	D_DEC.pth
M1	Multi-task	BIN / VQ / AUX / DEC	vq / aux / dec	M1_VQ_bal.pth, M1_AUX_bal.pth, M1_DEC_bal.pth
M2	Multi-task	VQ / AUX / DEC	vq / aux / dec	M2_VQ_bal.pth, M2_AUX_bal.pth, M2_DEC_bal.pth

SAST Net

Files are in model_checkpoints/SAST_Net/.

Task	Trained on	Files
BIN	vq / aux / dec	SAST_Net_BIN_VQ_bal.pth, SAST_Net_BIN_AUX_bal.pth, SAST_Net_BIN_DEC_bal.pth
VQ	vq	SAST_Net_VQ.pth
AUX	aux	SAST_Net_AUX.pth
DEC	dec	SAST_Net_DEC.pth

Pretrained weights

Files are in model_checkpoints/Pretrain_weight/.

File	Description	Used by
xlsr2_300m.pt	Wav2Vec2 pretrained weight	Wav2Vec2-AASIST, SAST Net
mae_pretrained_base.pth	AudioMAE pretrained on AudioSet	SAST Net
tuned_weight.pth	Wav2Vec2-AASIST on CodecFake+	SAST Net

Usage

Download a checkpoint:

from huggingface_hub import hf_hub_download

checkpoint = hf_hub_download(
    repo_id="CodecFake/CodecFake_Source_Tracing",
    filename="model_checkpoints/SAST_Net/SAST_Net_VQ.pth",
)

Setup, inference, and training commands are in the GitHub README.

Note: The metadata/ files in the GitHub repository use the file names from our internal copy of the dataset. The public CodecFake+ release uses a different layout and different CoSG file names, so the paths need to be adapted before running the scripts.

Citation

@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}
}

@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}
}

@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}
}