W2V-large-tidylang

Overview

This model is the official baseline system for the TidyLang2026 Language Recognition Challenge.

🔗 Challenge website: https://tidylang2026.github.io/

The model is trained for language recognition and serves as the reference system for evaluation and benchmarking.

---

Model Description

• Architecture: Wav2Vec2-Large backbone
• Task: Language Recognition
• Training Data: TidyVoice multilingual dataset
• Output: language embeddings

This model extracts fixed-dimensional language embeddings that can be used for:

• language verification (EER, minDCF)
• language identification
• Cross-lingual speaker analysis

---

Evaluation

The official evaluation pipeline is available here:

🔗 https://github.com/areffarhadi/TidyLang2026-baseline

Please use the provided evaluation scripts to reproduce challenge metrics.

---

Citation

If you use this model, please cite:

@misc{farhadi2026tidy,
      title={TidyVoice: A Curated Multilingual Dataset for Speaker Verification Derived from Common Voice},
      author={Aref Farhadipour and Jan Marquenie and Srikanth Madikeri and Eleanor Chodroff},
      year={2026},
      journal={ICASSP2026},
      url={https://arxiv.org/abs/2601.16358},
}