VoxTubeS is a speaker-anonymized synthetic derivative of VoxTube, released under CC BY-NC-SA 4.0. By requesting access you agree to: (1) use the data for non-commercial research only; (2) attribute VoxTubeS and VoxTube (Yakovlev et al., 2023); (3) license any derivatives under CC BY-NC-SA 4.0; and (4) not attempt to re-identify or de-anonymize the underlying speakers, nor use the audio to impersonate any individual. The synthetic voices do not correspond to real people.

VoxTubeS

VoxTubeS is a family of speaker-anonymized synthetic speech corpora derived from the VoxTube corpus (distributed by ID R&D under CC BY-NC-SA 4.0), released with the SLT 2026 paper "VoxTubeS: Distributable Speaker-Anonymized Synthetic Speech Corpora and Their Analysis."

Each config is an utterance-aligned English subset of 1,292,167 utterances (1,163,387 train + 128,780 dev) from 1,511 speakers, 16 kHz mono. Utterance IDs and speaker labels are identical across every config, so the variants are paired for controlled comparison.

Configs (operating points)

Config	Method	Notes
`ohnn-hifigan`	OHNN embedding-transform anonymization + HiFiGAN vocoder	environment remixed
`ohnn-bigvgan-sc`	OHNN anonymization + BigVGAN vocoder with a speaker-consistency (SC) loss	environment remixed
`salt-k4`	SALT latent kNN voice conversion, 4-speaker mixture (k=4)	environment remixed
`salt-k8`	SALT latent kNN voice conversion, 8-speaker mixture (k=8)	environment remixed
`daien-ncfg-m1.0`	DAIEN-TTS with negative speaker-prompt guidance, γ = −1.0	strongest identity suppression
`daien-ncfg-m0.75`	DAIEN-TTS, γ = −0.75	intermediate
`daien-ncfg-m0.5`	DAIEN-TTS, γ = −0.5	mildest suppression
`auth`	authentic reference (metadata only)	maps IDs → public VoxTube provenance

auth uses the same WebDataset container as the audio configs so it loads identically, but its samples are metadata only — the original un-anonymized audio is not redistributed. Reconstruct it from the public VoxTube corpus using the video_id / segment_index fields.

Format (unified)

Every config is a WebDataset of uncompressed tar shards. Each sample shares one key scheme (<speaker>-<video>-seg<index>) and one metadata schema.

Audio configs (7) — one uniform audio format: FLAC, 16 kHz, mono, PCM_16. Clip durations are the pipelines' natural lengths (not padded/trimmed; ≈2.0–4.0 s). Each sample is <ID>.flac + <ID>.json:
```
{"id": "<speaker>-<video>-seg<index>", "spk_id": "<speaker>",
 "split": "train", "variant": "OHNN-BigVGAN-SC",
 "samplerate": 16000, "num_frames": 64000, "duration": 4.0}
```

auth config — same container, <ID>.json only (no .flac):

{"id": "<speaker>-<video>-seg<index>", "spk_id": "<speaker>",
 "video_id": "<video>", "segment_index": 4, "split": "train",
 "variant": "Auth", "source": "VoxTube", "audio": "not_redistributed"}

Usage

from datasets import load_dataset

# Stream one variant (no full download)
ds = load_dataset("nii-yamagishilab/VoxTubeS", "ohnn-bigvgan-sc",
                  split="train", streaming=True)
ex = next(iter(ds))
audio = ex["flac"]          # decoded array + sampling_rate
meta  = ex["json"]          # id, spk_id, split, variant, ...

# The un-anonymized reference loads the same way but carries metadata only:
auth = load_dataset("nii-yamagishilab/VoxTubeS", "auth", split="train")
prov = next(iter(auth))["json"]     # video_id, segment_index -> fetch from VoxTube

Citation

If you use VoxTubeS, please cite the paper. You may also cite the dataset record by its DOI.

Paper:

@inproceedings{voxtubes2026,
  title     = {VoxTubeS: Distributable Speaker-Anonymized Synthetic Speech Corpora and Their Analysis},
  author    = {Zhang, Zhe and Lu, Yexin and Yamagishi, Junichi},
  booktitle = {Proc. IEEE Spoken Language Technology Workshop (SLT)},
  year      = {2026}
}

Dataset record — 10.5281/zenodo.22699494 (datasheet, manifests, evaluation tables, checksums):

@dataset{voxtubes_data_2026,
  title     = {VoxTubeS: Distributable Speaker-Anonymized Synthetic Speech Corpora},
  author    = {Zhang, Zhe and Lu, Yexin and Yamagishi, Junichi},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22699494}
}

License & attribution

CC BY-NC-SA 4.0, inherited from VoxTube (ID R&D). This means: attribution required, non-commercial use only, and derivatives must be shared under the same license. You must credit both VoxTubeS and VoxTube (Yakovlev et al., 2023); the SALT variants also use LibriSpeech (CC BY 4.0) target speakers. Do not use this data to re-identify VoxTube speakers or to impersonate real individuals — the synthetic voices are not real people. See LICENSE.txt and LICENSES-THIRD-PARTY.md in the repo, and the Zenodo record for the full datasheet, evaluation results, and checksums.

Downloads last month: 22

Total file size:

672 GB