README.md

metadata

configs:
  - config_name: en
    data_files:
      - path: en/metadata.jsonl
        split: test
  - config_name: de
    data_files:
      - path: de/metadata.jsonl
        split: test
  - config_name: es
    data_files:
      - path: es/metadata.jsonl
        split: test
  - config_name: fr
    data_files:
      - path: fr/metadata.jsonl
        split: test
  - config_name: it
    data_files:
      - path: it/metadata.jsonl
        split: test
language:
  - de
  - en
  - es
  - fr
  - it
pretty_name: RedVox
size_categories:
  - 1K<n<10K
tags:
  - safety
  - fairness
  - stereotypes
  - multilingual
  - speech
  - audio
extra_gated_prompt: >-
  You must agree not to use the dataset as training data, for voice cloning, or
  for identification purposes.
extra_gated_fields:
  I want to use this dataset for:
    type: select
    options:
      - Research
      - Education
      - label: Other
        value: other
  I agree not to use this dataset for training: checkbox
  I agree not to use this dataset for voice cloning or speaker identification: checkbox

RedVox

Multilingual red teaming dataset for audio and speech.

This dataset corresponds to the test set presented in the paper "RedVox: Safety and Fairness Gaps in Speech Models Across Languages" (Savoldi, Papi et al., 2026)

Dataset Structure

The dataset is organized by language configuration:

• en/ - English language samples (1,359 entries)
• de/ - German language samples (519 entries)
• es/ - Spanish language samples (354 entries)
• fr/ - French language samples (401 entries)
• it/ - Italian language samples (781 entries)

Total: 3,414 entries

Each directory contains:

• metadata.jsonl - JSONL file with sample metadata and annotations
• audio/ - Directory with audio files in WAV format

Audio Modalities

The dataset includes two main types of input types:

Speech Modality

• Original speech recordings containing harmful requests uttered by human participants
• Single entry per sample
• audio_type: "speech"

Audio Modality

• Available into 3 audio variants
  • audio_type: "silence" - Background silence
  • audio_type: "noise_a" - Background noise variant A
  • audio_type: "noise_b" - Background noise variant B

Metadata Schema

{
  "ID": 1,
  "audio": "audio/filename.wav",
  "original_id": 0,
  "user_id": "hash_string",
  "lang": "de",
  "vulnerability_type": "stereotype",
  "audio_type": "speech",
  "modality": "speech",
  "user_text": "text input or question",
  "transcript": "transcribed speech content"
}

Features

• ID: unique entry identifier
• audio: Path to input audio file
• user_id: Unique identifier for user
• lang: Source language
• original_id: original dataset identifier from M-ALERT and SHADES datasets
• vulnerability_type: Type of vulnerability being tested (stereotype or unsafe request)
• attack_type: Type of attack (non_adversarial)
• audio_type: Type of audio (speech / silence / noise_a / noise_b)
• user_text: User's text input
• modality: harmful content modality (in speech or text)
• gender: User's gender
• age_group: User's age group
• mother_tongue: User's mother tongue
• ethnicity: User's ethnicity
• transcript: Speech transcript of the audio input

Stats by Language

lang	total	speech	silence	noise_a	noise_b
de	519	135	128	128	128
en	1359	342	339	339	339
es	354	85	89	90	90
fr	401	95	102	102	102
it	781	184	199	199	199

Splits

The dataset can be loaded with different configurations:

from datasets import load_dataset

# Load German language config
dataset = load_dataset("path/to/repo", name="de", split="test")

# Load English language config
dataset = load_dataset("path/to/repo", name="en", split="test")

Bias, Risks, and Limitations

• Consent: All voice recordings and associated participant metadata in this dataset come from individuals who gave explicit, informed consent for their data to be published and used for research purposes.
• Harmful content by design: RedVox is a red-teaming dataset and is intended solely for safety and robustness evaluation, not as representative or endorsed text/speech.
• Speaker privacy and re-identification risk: Users must not attempt to re-identify, or deanonymize any speaker in this dataset, whether through manual inspection, automated techniques, or cross-referencing with other data sources.

Citation

If you use this dataset in your research, please cite the following paper:

@misc{savoldi2026redvoxsafetyfairnessgaps,
      title={RedVox: Safety and Fairness Gaps in Speech Models Across Languages}, 
      author={Beatrice Savoldi and Sara Papi and Wafa Aissa and Matteo Negri and Luisa Bentivogli},
      year={2026},
      eprint={2606.26968},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.26968}
}

Code and Evaluation

For code, evaluation scripts, and additional resources related to this dataset, see: hlt-mt/redvox

License

RedVox is derived from two source datasets, and each subset is released under the same license as its original source.

• SHADES subset (vulnerability_type: stereotype) — derived from SHADES and released under the same Montreal Data License as the original. Use requires acceptance of the Acceptable Use Policy in that license (no harmful, discriminatory, or law-enforcement/immigration use; no use to identify further training data; distribution only under the same terms; attribution to SHADES required).

• M-ALERT subset (all other entries, vulnerability_type: unsafe_request*`) — derived from M-ALERT and released under the same CC BY-NC-SA 4.0 license as the original (non-commercial use, attribution required, share-alike).

This repository is gated. Requesting access requires agreeing to use RedVox exclusively for non-commercial AI safety and fairness research, and to comply with the Acceptable Use Policy.