VoxAnon

A balanced dataset for a fair evaluation of speech anonymization.

• 📄 Paper: VoxAnon: A balanced dataset for a fair evaluation of speech anonymization
• 💻 Code, speaker lists, utterance lists and speaker attributes: GitHub repository URL

This repository contains the balanced configuration of VoxAnon: the original audio and its anonymized versions produced by the baseline anonymizers of the 2026 VoicePrivacy Challenge:

Version	System
original	Original (non-anonymized) audio
B3	DigitalPhonetics/speaker-anonymization
B4	m-pana/spk_anon_nac_lm
B5	deep-privacy/SA-toolkit

Dataset structure

Audio is stored as .wav files packed into .tar shards, one group of shards per version and per split:

VoxAnon/
├── orig/
│   ├── train/
│   │   ├── shard-000000.tar
│   │   └── ...
│   └── test/
│       ├── shard-000000.tar
│       └── ...
├── B3/   (same layout)
├── B4/   (same layout)
└── B5/   (same layout)

Each shard contains .wav utterances. Speaker attributes (gender, age, accent, pitch level) and the lists of speakers and utterances are provided in the GitHub repository.

Download

Option 1: download the raw shards

pip install -U huggingface_hub
hf download OraneDufour/VoxAnon --repo-type dataset --local-dir ./VoxAnon

Or in Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="OraneDufour/VoxAnon",
    repo_type="dataset",
    local_dir="./VoxAnon",
)

To download only some versions or splits, filter with --include (CLI) or allow_patterns (Python):

hf download OraneDufour/VoxAnon --repo-type dataset --local-dir ./VoxAnon --include "B4/*.tar"

snapshot_download(
    repo_id="OraneDufour/VoxAnon",
    repo_type="dataset",
    local_dir="./VoxAnon",
    allow_patterns=["original/*.tar", "B4/*.tar"],
)

Once downloaded, shards are standard tar archives:

tar -xf VoxAnon/B4/test/shard-000000.tar -C ./B4_test

Option 2: stream with 🤗 datasets

from datasets import load_dataset

ds = load_dataset("OraneDufour/VoxAnon", "B4", split="test", streaming=True)

for sample in ds:
    print(sample.keys())
    break

Without streaming=True, the data is downloaded and converted to a local cache. Requires pip install datasets (plus an audio backend such as soundfile or torchcodec to decode the .wav files).

Citation

@misc{dufour2026voxanonbalanceddatasetfair,
      title={VoxAnon: A balanced dataset for a fair evaluation of speech anonymization}, 
      author={Orane Dufour and Paul Magron and Mickael Rouvier},
      year={2026},
      eprint={2610.13313},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2610.13313}, 
}