Loading datasets...
License:
CC0-1.0
Steward:
TidyVoice2026 ChallengeDataset ID:
cmihtsewu023so207xot1iqqw
Release Date: 11/27/2025
Format: WAV
Size: 36.72 GB
Share
This dataset is designed for speaker verification using the Mozilla Common Voice corpus across 40 languages. It includes approximately 5,000 speakers who each have recordings in more than one language. Leveraging this multilingual overlap, we construct the trial pairs to explore cross-lingual variation in the speaker verification task.
Forbidden Usage
According to Mozilla’s usage rules, it is forbidden to use this dataset for speaker identification or for recovering a speaker’s identity. The data MUST only be used for speaker verification tasks.
Intended Use
All rules and restrictions are the same as those of the original Mozilla Common Voice datasets.
TidyVoiceX is a large-scale, multilingual speech corpus specifically curated for cross-lingual speaker verification research. Derived from Mozilla Common Voice (MCV), this dataset is designed to isolate the effect of language switching across multiple languages, enabling focused investigation into language-independent speaker embeddings.
This dataset is part of the TidyVoice2026 Challenge, an official challenge at Interspeech 2026 focused on advancing cross-lingual speaker verification systems.
🌐 Challenge Website: https://tidyvoice2026.github.io/
| Metric | Training Set | Development Set | Total |
|---|---|---|---|
| Speakers | 3,666 | 808 | 4,474 |
| Languages | 40 | 40 | 40 |
| Utterances | 262,000 | 60,000 | 321,711 |
| Duration (hours) | 370 | 87 | 457 |
| Domain | Read Speech | Read Speech | Read Speech |
The Tidy-X dataset includes the following 40 languages exclusively for training:
ab (Abkhazian)
ar (Arabic)
ba (Bashkir)
be (Belarusian)
bg (Bulgarian)
bn (Bengali)
ca (Catalan)
cv (Chuvash)
cy (Welsh)
de (German)
dv (Dhivehi)
el (Greek)
en (English)
fa (Persian)
fr (French)
ha (Hausa)
hi (Hindi)
hsb (Upper Sorbian)
hy-AM (Armenian)
ja (Japanese)
ka (Georgian)
lg (Luganda)
lt (Lithuanian)
mk (Macedonian)
ml (Malayalam)
mr (Marathi)
nl (Dutch)
or (Odia)
pl (Polish)
pt (Portuguese)
ru (Russian)
ta (Tamil)
th (Thai)
tk (Turkmen)
tr (Turkish)
ug (Uyghur)
uz (Uzbek)
yo (Yoruba)
yue (Cantonese)
zh-CN (Chinese)
📥 Download: Development Set Trial Pairs
The trial pairs file contains the trial pairs for the development set, formatted according to the challenge specifications.
Multilingual Scope: 40 training languages covering diverse language families
Cross-lingual Focus: Designed to evaluate speaker verification under language mismatch conditions
Pseudonymized IDs: All speaker identities are pseudonymized to protect privacy
Controlled Domain: Read speech domain minimizes stylistic and phonetic variability
Open Access: Publicly available for research purposes
Standardized Splits: Clear train/development separation for reproducible research
Audio Format: WAV format, 16 kHz sampling frequency
This dataset is specifically designed for:
Cross-lingual speaker verification research
Language-independent speaker embedding development
Multilingual speaker recognition systems
Fairness and bias evaluation in speaker verification
If you use the Tidy-X dataset in your research, please cite:
@inproceedings{farhadipour2026tidyvoice,
title={TidyVoice Challenge: Cross-Lingual Speaker Verification},
author={Farhadipour, Aref and Marquenie, Jan and Madikeri, Srikanth and Vukovic, Teodora and Dellwo, Volker and Reid, Kathy and Tyers, Francis M. and Siegert, Ingo and Chodroff, Eleanor},
booktitle={Interspeech 2026},
year={2026}
}
This dataset is derived from Mozilla Common Voice. Please refer to the Mozilla Common Voice license for usage terms and conditions.
For questions, issues, or contributions, please visit:
Challenge Website: https://tidyvoice2026.github.io/
This dataset was created as part of the TidyVoice2026 Challenge at Interspeech 2026, developed by researchers from the University of Zurich, Otto-von-Guericke-University Magdeburg, Mozilla Foundation, Indiana University and Australian National University.