TidyVoiceM_ASV

TidyVoiceM_ASV Dataset

Overview

Tidy-M is a large-scale, massively multilingual speech corpus curated for intra-lingual (same-language) speaker verification research. It is the monolingual partition of the TidyVoice dataset, derived from Mozilla Common Voice (MCV) after mitigating the speaker heterogeneity that affects MCV client IDs.

🌐 Project / models: https://github.com/areffarhadi/wespeaker/tree/master/examples/tidyvocie

Dataset Statistics

Splits

Splits prioritize a large training set:

• Train: speakers with more than four utterances.

• Test: the remaining speakers, with the constraint that every language has at least five speakers in the test set.

• Speakers are disjoint between train and test.

Trial Pairs

The Tidy-M evaluation condition is intra-lingual: both sides of every trial are in the same language.

The full test trial list contains approximately 2.8 million pairs. Trials are also provided per language for all 81 languages, enabling fine-grained analysis of model performance across diverse linguistic contexts.

📥 Download: Tidy-M test trial pairs

Trial file format (WeSpeaker/VoxCeleb style):

  
1 id010001/en/utt_0001.wav id010001/en/utt_0007.wav
0 id010001/en/utt_0001.wav id010042/en/utt_0003.wav

Key Features

• Massive speaker diversity: 212,617 speakers — larger than any prior public read-speech speaker-verification corpus

• Broad language coverage: 81 languages spanning many language families

• Curated identities: client IDs cleaned with a ResNet-293 verification pass (cosine threshold 0.4) to remove speaker-heterogeneous IDs

• Intra-lingual condition: target and non-target trials matched in language, isolating speaker discrimination from language effects

• Per-language trials: trial lists released per language for fine-grained, fair evaluation

• Pseudonymized IDs: all speaker identities are pseudonymized to protect privacy

• Controlled domain: read speech minimizes stylistic variability

• Standardized splits: speaker-disjoint train/test split for reproducible research

• Audio format: WAV, 16 kHz

Use Cases

• Large-scale multilingual speaker verification training and evaluation

• Pre-training / fine-tuning speaker embedding extractors

• Cross-domain generalization studies (read speech → conversational speech)

• Per-language fairness and bias evaluation of speaker verification systems

• Speaker-ID curation and data-quality research on crowdsourced corpora

• Anti-spoofing and speaker anonymization research

Citation

If you use the Tidy-M dataset in your research, please cite:

@inproceedings{farhadipour2026tidyvoice,
  title={TidyVoice: A Curated Multilingual Dataset for Speaker Verification Derived from Common Voice},
  author={Farhadipour, Aref and Marquenie, Jan and Madikeri, Srikanth and Chodroff, Eleanor},
  booktitle={ICASSP},
  year={2026}
}

@article{zhang2025quantifying,
  title={Quantifying and reducing speaker heterogeneity within the Common Voice corpus for phonetic analysis},
  author={Zhang, Miao and Farhadipour, Aref and Baker, Annie and Ma, Jiachen and Pricop, Bogdan and Chodroff, Eleanor},
  journal={arXiv preprint arXiv:2506.00733},
  year={2025}
}

License

This dataset is derived from Mozilla Common Voice and is distributed under the same terms (CC-0). Please refer to the Mozilla Common Voice license for usage terms and conditions.

Contact

• Project website: https://tidyvoice2026.github.io/

• Email: [email protected], [email protected]