HIR-SDD

HIR-SDD (Human-annotated Inference Reasoning for Spoofing and Deepfake Detection) is a dataset of speech clips with human-written natural-language reasoning and structured spoofing cue tags.

Each row links an audio clip to an explanation of why the clip was judged genuine or spoofed, together with a discrete set of perceptual reason tags.

• Paper: Towards Robust Speech Deepfake Detection via Human-Inspired Reasoning
• Dataset: marsianin500/HIR-SDD
• Raw annotations: marsianin500/HIR-SDD-raw

Dataset summary

Item	Value
Annotation rows	88,630
Unique audio clips	45,036
Crowd annotation rows	84,678
SpeechLLM-derived annotation rows	3,952

Task

Given a speech clip, models may be trained or evaluated to:

1. classify the clip as genuine (bonafide) or spoofed (fake);
2. predict structured reason tags;
3. generate or evaluate natural-language reasoning about the decision.

Annotation format

Each row is one annotation for one clip. The same audio_id may appear in multiple rows when several annotators labeled the same clip.

Field	Description
audio_id	Stable clip identifier
path	Relative path to bundled audio in this repository
source_corpus	Upstream audio corpus
source	Annotation origin: modern-generators or speechllm
reasoning	Natural-language explanation
reasons	List of structured spoofing cue tags
is_bonafide	Annotator label
language	Clip language metadata
marker_id	Annotator identifier (modern-generators rows)
consistency	Annotator consistency score (modern-generators rows)
annotator_quality	Annotator quality tier (modern-generators rows)
complexity	Clip complexity score (modern-generators rows)

Rows with source = speechllm contain reasoning, reasons, and is_bonafide only from the SpeechLLM annotation pipeline. Fields such as marker_id, consistency, annotator_quality, and complexity are empty for those rows.

Audio files

Audio is stored once per audio_id and referenced through path:

audio/
├── by_nc_sa/{shard}/{audio_id}.{ext}
└── by_sa/{shard}/{audio_id}.{ext}

{shard} is the first two characters of audio_id (keeps each directory under Hugging Face's per-folder file limit).

Approximate clip counts by upstream corpus:

source_corpus	Unique clips
final_dataset / ESpeech	7,097
ASVspoof5	6,484
LibriSeVoc	6,190
MLAAD	4,904
XTTS	4,471
SpeechLLM	3,952
Golos	3,935
DFADD	3,493
SOVA	1,596
Espeech_spoofs	1,445
M-AILABS	885
ruLS	584

Usage

from datasets import load_dataset

ds = load_dataset("marsianin500/HIR-SDD", split="train")
example = ds[0]
print(example["audio_id"], example["path"])
print(example["reasoning"])

License

HIR-SDD annotations are released under CC BY-NC-SA 4.0.

Bundled audio is redistributed under the terms of the original corpora and is split into two folders:

Folder	License	Corpora
audio/by_nc_sa/	CC BY-NC-SA 4.0	ASVspoof5, MLAAD, Golos, XTTS outputs, DFADD, SOVA, ruLS, M-AILABS, ESpeech, SpeechLLM
audio/by_sa/	CC BY-SA 4.0	LibriSeVoc

Some upstream sources impose additional terms:

• Golos: SberDevices license
• XTTS outputs: Coqui Public Model License
• ASVspoof5: see the official release terms

See also audio/by_nc_sa/LICENSE.md and audio/by_sa/LICENSE.md in the repository.

Audio sources

ASVspoof 5

@article{wang2025asvspoof5,
  title   = {{ASVspoof} 5: Design, Collection and Validation of Resources for Spoofing, Deepfake, and Adversarial Attack Detection Using Crowdsourced Speech},
  author  = {Wang, Xin and Delgado, H{\'e}ctor and Tak, Hemlata and Jung, Jee-weon and Shim, Hye-jin and Todisco, Massimiliano and Kukanov, Ivan and Liu, Xuechen and Sahidullah, Md and Kinnunen, Tomi and Evans, Nicholas and Lee, Kong Aik and Yamagishi, Junichi and others},
  journal = {arXiv preprint arXiv:2502.08857},
  year    = {2025}
}

MLAAD

@article{muller2024mlaad,
  title   = {{MLAAD}: The Multi-Language Audio Anti-Spoofing Dataset},
  author  = {M{\"u}ller, Nicolas M. and Kawa, Piotr and Choong, Wei Herng and Casanova, Edresson and G{\"o}lge, Eren and M{\"u}ller, Thorsten and Syga, Piotr and Sperl, Philip and B{\"o}ttinger, Konstantin},
  journal = {arXiv preprint arXiv:2401.09512},
  year    = {2024}
}

LibriSeVoc

@inproceedings{sun2023librisevoc,
  title         = {AI-Synthesized Voice Detection Using Neural Vocoder Artifacts},
  author        = {Sun, Chengzhe and Jia, Shan and Hou, Shuwei and Lyu, Siwei},
  booktitle     = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
  year          = {2023},
  eprint        = {2304.13085},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD}
}

Golos

@inproceedings{karpov2021golos,
  title     = {{Golos}: Russian Dataset for Speech Research},
  author    = {Karpov, Nikolay and Denisenko, Alexander and Minkin, Fedor},
  booktitle = {Interspeech 2021},
  pages     = {1419--1423},
  year      = {2021},
  doi       = {10.21437/Interspeech.2021-462}
}

M-AILABS Speech Dataset

@misc{mailabs2017,
  author       = {Solak, I. Celeste Aurora and Naumov, Dima},
  title        = {The {M-AILABS} Speech Dataset},
  year         = {2017},
  howpublished = {\url{https://github.com/i-celeste-aurora/m-ailabs-dataset}}
}

DFADD

@inproceedings{du2024dfadd,
  title        = {{DFADD}: The Diffusion and Flow-Matching Based Audio Deepfake Dataset},
  author       = {Du, Jiawei and Lin, I-Ming and Chiu, I-Hsiang and Chen, Xuanjun and Wu, Haibin and Ren, Wenze and Tsao, Yu and Lee, Hung-yi and Jang, Jyh-Shing Roger},
  booktitle    = {2024 IEEE Spoken Language Technology Workshop (SLT)},
  pages        = {921--928},
  year         = {2024},
  organization = {IEEE}
}

SOVA

@misc{sova2021dataset,
  author       = {Zubarev, Egor and Moskalets, Timofey and {SOVA.ai}},
  title        = {{SOVA} Dataset: Free Public {STT}/{ASR} Dataset},
  year         = {2021},
  howpublished = {\url{https://github.com/sovaai/sova-dataset}}
}

Russian LibriSpeech (ruLS)

@techreport{bakhturina2021toolbox,
  title       = {A Toolbox for Construction and Analysis of Speech Datasets},
  author      = {Bakhturina, Evelina and Lavrukhin, Vitaly and Ginsburg, Boris},
  year        = {2021},
  institution = {NVIDIA}
}

ESpeech

@dataset{espeech_podcasts2025,
  title        = {{ESpeech}: Large-Scale Russian Podcast Speech Dataset},
  author       = {{ESpeech Team}},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/ESpeech/ESpeech-podcasts}}
}

XTTS

@inproceedings{casanova24_interspeech,
  title     = {{XTTS}: a Massively Multilingual Zero-Shot Text-to-Speech Model},
  author    = {Casanova, Edresson and Davis, Kelly and G{\"o}lge, Eren and G{\"o}knar, G{\"o}rkem and Gulea, Iulian and Hart, Logan and Aljafari, Aya and Meyer, Joshua and Morais, Reuben and Olayemi, Samuel and Weber, Julian},
  booktitle = {Interspeech 2024},
  pages     = {4978--4982},
  year      = {2024},
  doi       = {10.21437/Interspeech.2024-2016}
}

SpeechEval / SpeechLLM

@article{wang2025speechllm,
  title   = {{SpeechLLM}-as-Judges: Towards General and Interpretable Speech Quality Evaluation},
  author  = {Wang, Hui and Zhao, Jinghua and Yang, Yifan and Liu, Shujie and Chen, Junyang and Zhang, Yanzhe and Zhao, Shiwan and Li, Jinyu and Zhou, Jiaming and Sun, Haoqin and others},
  journal = {arXiv preprint arXiv:2510.14664},
  year    = {2025}
}

Citing HIR-SDD

If you use this dataset or the HIR-SDD framework, please cite:

@misc{dvirniak2026robustspeechdeepfakedetection,
  title         = {Towards Robust Speech Deepfake Detection via Human-Inspired Reasoning},
  author        = {Dvirniak, Artem and Kushnir, Evgeny and Tarasov, Dmitrii and Iudin, Artem and Kiriukhin, Oleg and Pautov, Mikhail and Korzh, Dmitrii and Rogov, Oleg Y.},
  year          = {2026},
  eprint        = {2603.10725},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2603.10725}
}

You may also reference the Hugging Face release:

@misc{hir_sdd_hf_2026,
  title        = {{HIR-SDD}: Human-annotated Inference Reasoning for Spoofing and Deepfake Detection},
  author       = {Dvirniak, Artem and Kushnir, Evgeny and Tarasov, Dmitrii and Iudin, Artem and Kiriukhin, Oleg and Pautov, Mikhail and Korzh, Dmitrii and Rogov, Oleg Y.},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/marsianin500/HIR-SDD}
}