🦙 LlamaPartialSpoof

📄 Preprint | 📢 Speech Samples | ⬇️ Download

This is the official repository of the LlamaPartialSpoof Project. The dataset is ready for download from zenodo. Stay updated on the latest resources and information about the LlamaPartialSpoof project and partially fake speech by following this repository.

Updates

• 2025-01-07 The paper "LlamaPartialSpoof: An LLM-Driven Fake Speech Dataset Simulating Disinformation Generation" is accepted, we will give a presentation on this topic at ICASSP 2025 in Hyderabad, India around April 6 to 11, 2025

Related Repositories

• MultiResoModel (Simple)
• partialspoof-metrics

FAQ

01. How should the data be used?

• LlamaPartialSpoof was designed for evaluation hence we didn't provided training data. We want researchers to develop a system that either trained on a third-party dataset (e.g. PartialSpoof) or using a non-training method so it can generalize to real life scenarios.
• In the paper Robust Localization of Partially Fake Speech: Metrics, Models, and Out-of-Domain Evaluation we split the dataset speakers into train and test subset to test continous learning scenarios. You can find the data partitions in the split/ directory.

02. Baseline mode

• I prepared a baseline system with the processing pipeline to benchmark this dataset at MultiResoModel (Simple)

03. Modified transcipts

• The original dev-clean, test-clean transcripts and the modified dev01 transcripts are included in this repository

Have a question?

If you have any question or feedback, feel free to create a new issue in this repository. It may take sometimes for us to get back but we love to connect with the research community on this topic.

Citations

• The original LlamaPartialSpoof paper

@inproceedings{luong2025llamapartialspoof,
  title={LlamaPartialSpoof: An LLM-Driven Fake Speech Dataset Simulating Disinformation Generation},
  author={Luong, Hieu-Thi and Li, Haoyang and Zhang, Lin and Lee, Kong Aik and Chng, Eng Siong},
  booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2025},
  organization={IEEE}
}

• The extended analysis on the localization task

@article{luong2025robust,
  title={Robust Localization of Partially Fake Speech: Metrics, Models, and Out-of-Domain Evaluation},
  author={Luong, Hieu-Thi and Rimon, Inbal and Permuter, Haim and Lee, Kong Aik and Chng, Eng Siong},
  journal={arXiv preprint arXiv:2507.03468},
  year={2025}
}