Datasets:
Request access to GenTraceBench
Submit an access request. The dataset maintainers will review it manually; submitting this form does not grant immediate access to the data files.
Please provide your affiliation and intended use. If you are not affiliated with an institution, enter Independent researcher. Your Hugging Face username, email address, and form responses will be shared with the dataset maintainers to review and manage your access request. Approval is required to download gated files and does not replace the applicable license or third-party permissions.
Log in or Sign Up to review the conditions and access this dataset content.
GenTraceBench
GenTraceBench: A Benchmark for Tracing Audio Deepfakes Across Pre- and Post-training Stages — accepted at ISCSLP 2026.
GenTraceBench studies whether forensic fingerprints survive changes to a speech generator, including supervised fine-tuning, preference optimization, and pre-training-data composition. It covers 5 generator families, 16 conditions, and 49,728 synthetic utterances, with matched source texts and speaker prompts across conditions.
The paper evaluates binary detection, closed-set source attribution, and open-set verification under a train-on-foundation, test-on-adapted protocol. This repository is named GenTrace; the benchmark and paper are named GenTraceBench.
Access requests
Sign in to Hugging Face, provide your affiliation and intended use, and submit an access request. Requests require manual approval. The dataset description remains public; gated files require an approved account and authenticated downloads. Maintainers and authorized repository collaborators may retain direct access.
Hugging Face shares the applicant's username, email address, and form responses with the maintainers for access review and administration. Please do not include confidential information in the form. Approval does not replace applicable licenses or third-party permissions.
What is included
- Original synthetic WAV files, packaged into one TAR archive per variant; no release-time resampling or lossy conversion.
- The exact original train/validation/test prompt lists and five-way attribution labels.
- Full sample metadata, checksums, actual paper-use partitions, and a dependency-free extraction/verification helper.
- English and Chinese download and usage instructions.
Not included: Seed-TTS bona fide/reference recordings, source transcripts, generator weights, or trained forensic checkpoints. Binary detection requires obtaining the real-speech counterparts separately; see source notes.
Quick start
After your access request is approved:
pip install -U huggingface_hub
hf auth login
hf download wli3221134/GenTrace --repo-type dataset --local-dir GenTraceBench
python GenTraceBench/scripts/gentrace.py extract --root GenTraceBench
python GenTraceBench/scripts/gentrace.py stats --root GenTraceBench
Extraction creates GenTraceBench/audio/ with the original relative model paths. For selective downloads, Python loading examples, all three tasks, and checksum verification, see USAGE.md or 中文用法.
Variants
| Family | Conditions | Count |
|---|---|---|
| CosyVoice2 | Pre-trained; DPO; SFT on NVSpeech | 3 |
| F5-TTS | Pre-trained; DPO | 2 |
| FlexiVoice | Pre-trained; S1; S1+S2; S1+S2+S3 | 4 |
| MaskGCT | Pre-trained; DPO | 2 |
| Vevo2 | Emilia; DPO; GRPO; SingNet; Emilia+SingNet | 5 |
Every variant contains 3,108 matched examples: 1,088 English and 2,020 Chinese. metadata/variants.json gives the exact directory-to-paper mapping; metadata/release.json contains measured release statistics.
The original audio is 24 kHz, mono, totaling approximately 73.97 hours and 16.06 GB before TAR overhead. All 49,728 files passed decoding and finite-sample checks; no identical decoded-audio duplicates were found in this release.
Preserve the evaluation protocol
The source prompt split contains 1,035 training, 517 validation, and 1,556 test examples. It is shared across all variants. The entire dataset is not a test set.
| Actual paper use | Synthetic samples |
|---|---|
| Training: five foundation variants | 5,175 |
| Validation: five foundation variants | 2,585 |
| Testing: all 16 variants on held-out prompts | 24,896 |
| Additional parallel samples, unused in the paper's protocol | 17,072 |
Use protocols/train.csv, protocols/validation.csv, and protocols/test.csv. Do not accidentally train on the held-out adaptation conditions. prompt_split describes the original matched prompt partition; protocol_split describes actual use by the paper protocol.
Exact historical verification trial lists are not part of this release. Newly sampled pairs are not the paper's original trials; record your trial list, grouping rules, and random seed. See usage for the distinction between family labels and variant labels.
Limitations and sources
The benchmark covers controlled, clean Chinese/English audio from five families, not all deployment conditions. Synthesis quality varies, particularly for the SingNet-only control. The release checks prompt-level partition integrity; it does not claim speaker-disjoint partitions or anonymous voice identities.
Prompts originate from Seed-TTS Eval. See SOURCE_NOTES.md for upstream materials, licensing boundaries, and real-speech reconstruction guidance. The existing repository license metadata is retained; access approval does not confer additional rights to third-party materials.
Citation
@inproceedings{wang2026gentracebench,
title = {GenTraceBench: A Benchmark for Tracing Audio Deepfakes Across Pre- and Post-training Stages},
author = {Li Wang and Kunyu Feng and Wan Lin and Dekun Chen and Qinke Ni and Xueyao Zhang and Lei Wang and Jie Shi and Haizhou Li and Zhizheng Wu},
booktitle = {ISCSLP},
year = {2026}
}
For questions or correction requests, contact liwang1@link.cuhk.edu.cn or wuzhizheng@cuhk.edu.cn.
- Downloads last month
- 30