ODSS — An Open Dataset of Synthetic Speech

Benchmark-ready packaging of ODSS (An Open Dataset of Synthetic Speech), a multilingual (English / German / Spanish), multispeaker dataset for synthetic speech detection: each natural utterance is paired with TTS re-synthesis of the same text.

Overview

Binary classification: bonafide (natural human recordings) vs. spoof (text-to-speech). The synthetic side is generated by two TTS systems — the end-to-end VITS architecture and a two-step FastPitch + HiFi-GAN pipeline — over 156 voices spanning three languages with balanced gender. Label is the top-level generator directory: natural/ → bonafide, vits/ and fastpitch-hifigan/ → spoof.

Source corpora — VCTK-derived and others

ODSS draws its speech from four public corpora (this is a VCTK-based dataset among others):

Corpus	Language	Role
VCTK	English	English VITS voices (vits/vctk)
Hi-Fi TTS	English	natural + VITS + FastPitch
HUI-audio-corpus-german	German	natural + VITS + FastPitch
OpenSLR Spanish	Spanish	natural + VITS + FastPitch

Note: the vits/vctk synthetic clips (English) are present on the spoof side; their natural VCTK counterparts are distributed separately by the VCTK project and are not part of this release.

License & redistribution

Released under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license — the license shipped in the upstream ODSS LICENSE file. Redistribution and adaptation are permitted with attribution and ShareAlike (this packaging is itself CC BY-SA 4.0). Audio is the original 16 kHz mono WAV, embedded bit-exactly (no re-encode — a full decode probe of all 26,954 clips passed cleanly). See LICENSE.txt. Attribute the ODSS authors and the four source corpora.

Schema

Column	Type	Description
path	string	source-relative path, e.g. vits/vctk/p293/p293_168.wav, unique
audio	Audio(16000)	16 kHz mono WAV
label	ClassLabel	"bonafide" (0) / "spoof" (1)
notes	string	JSON: utterance_id, generator, source_corpus, speaker, language, attack

notes example:

{"utterance_id": "vits__vctk__p293__p293_168", "generator": "vits", "source_corpus": "vctk", "speaker": "p293", "language": "en", "attack": "vits"}

utterance_id is the full source-relative path with / → __ — the bare stem repeats across the three generators (the same texts are synthesized), so the generator prefix is what makes ids unique.

Quick Start

from datasets import load_dataset

ds = load_dataset("SpeechAntiSpoofingBenchmarks/ODSS", split="test")
print(ds[0])

Stats

Stat	Value
Total trials	26,954
Bonafide (natural)	7,961
Spoof (TTS)	18,993
— VITS	11,032
— FastPitch + HiFi-GAN	7,961
Languages	en (14,405), de (5,778), es (6,771)
Source corpora	VCTK, Hi-Fi TTS, HUI-de, OpenSLR-ES

Source provenance

• Paper: An Open Dataset of Synthetic Speech, IEEE, 2023. https://ieeexplore.ieee.org/document/10374863/
• Dataset: https://zenodo.org/records/8370669 (DOI 10.5281/zenodo.8370669)
• Project: vera.ai / Fraunhofer IDMT.

Evaluation

For evaluation instructions and submission format, see submissions/README.md.

Citation

@inproceedings{odss2023,
  title     = {{An Open Dataset of Synthetic Speech}},
  booktitle = {Proc. IEEE Workshop},
  year      = {2023},
  doi       = {10.5281/zenodo.8370669},
  note      = {ODSS, https://ieeexplore.ieee.org/document/10374863/},
}

Maintainer

Contact: k.n.borodin@mtuci.ru