Elamur Research⌕
☰
home / papers / 2026 / CS-YODAS: A Mined Dataset of In-the-Wild Code-Switched Speech

CS-YODAS: A Mined Dataset of In-the-Wild Code-Switched Speech

Summary

Large multilingual speech corpora usually tag each utterance with one language and filter out mixed-language material. As a result, natural code-switching (CS) is scarce in training data, and existing CS speech sets tend to be small, tied to one domain, elicited, or synthetic. The authors build CS-YODAS, a Creative Commons resource of spontaneous code-switched speech taken from YouTube audio in the cleaned YODAS corpus used by OWSM v4.

Their mining pipeline has two stages:

  • Candidate detection: an LLM (Qwen3-14B) reads each transcript and lists the languages it finds. Lines with two or more languages become candidates.
  • Validation: bilingual volunteers answer five questions about sample segments. Their answers then serve as in-context examples so the LLM can screen every candidate the same way. This removes common false alarms such as transcript errors, proper nouns and cognates.

Validation lifts precision from 18% to 70%. The release keeps 15 seconds of surrounding audio around each switch so researchers can study the context in which switching happens. Baseline experiments on spoken language identification suggest synthetic CS data cannot yet replace real recordings.

Key points

  • Size and languages: about 313 hours (312.7 h) of contextual audio across 7 matrix languages: Arabic, Mandarin Chinese, Czech, French, Hindi, Japanese and Russian. It is available on Hugging Face as byan/cs-yodas.
  • Source corpus: the cleaned YODAS from OWSM v4, which holds 166k hours in 75 languages.
  • Precision gain: human-guided LLM validation raised precision from 18% (700 mine_iter0 samples) to 70% (120 mine_iter1 samples). Hindi reached 82.9%, Chinese 76.9% and Arabic 55.2%, while Japanese and Russian yielded few true positives.
  • Share of switching: only 37.3 h (11.9%) of the audio is intra-sentential CS. French supplies most of the data (273.7 h), and Hindi has the highest CS rate (29.7%).
  • LID baseline: accuracy on CS-FLEURS READ stays at zero until the model has more than 5 hours of CS-YODAS training data.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3fir5kp98qyt

@inproceedings{yan2026cs,
  title = {{CS-YODAS: A Mined Dataset of In-the-Wild Code-Switched Speech}},
  author = {Yan, Brian and Wang, Qingzheng and Wiesner, Matthew and Diwan, Anuj and Iakovenko, Olga and Polok, Alex and Hamed, Injy and Shimizu, Shuichiro and Emerman, Iris and Hain, Thomas and Mortensen, David R. and Viechnicki, Peter and Watanabe, Shinji},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/3fir5kp98qyt},
  url = {https://doi.org/10.63317/3fir5kp98qyt}
}

References

30 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saunders, Francis Tyers, and Gregor Weber. 2020. Common Voice: A massively-multilingual speech corpus. In Proc. LREC, pages 4218–4222. find our card
  2. Laurie Burchell, Alexandra Birch, Robert P Thompson, and Kenneth Heafield. 2024. Code-switched language identification is harder than you think. arXiv preprint arXiv:2402.01505. find
  3. Shammur Absar Chowdhury, Amir Hussein, Ahmed Abdelali, and Ahmed Ali. 2021. Towards one model to rule all: Multilingual strategy for dialectal code-switching Arabic ASR. In Proc. Interspeech, pages 2466–2470. find our card
  4. Alexis Conneau, Min Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Jason Riesa, Clara Rivera, and Ankur Bapna. 2023. FLEURS: Few-shot learning evaluation of universal representations of speech. In Proc. IEEE SLT, pages 798–805. find our card
  5. Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. 2019. ArcFace: Additive angular margin loss for deep face recognition. In Proc. CVPR. find
  6. Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck. 2020. ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification. In Proc. Interspeech, pages 3830–3834. find
  7. Margaret Deuchar, Peredur Davies, Jon Herring, M Carmen Parafita Couto, and Diana Carter. 2014. Building bilingual corpora. Advances in the Study of Bilingualism, pages 93–111. find
  8. Anuj Diwan, Rakesh Vaideeswaran, Sanket Shah, Ankita Singh, Srinivasa Raghavan, Shreya Khare, Vinit Unni, Saurabh Vyas, Akash Rajpuria, Chiranjeevi Yarra, Ashish Mittal, Prasanta Kumar Ghosh, Preethi Jyothi, Kalika Bali, Vivek Seshadri, Sunayana Sitaram, Samarth Bharadwaj, Jai Nanavati, Raoul Nanava… find
  9. MUCS 2021: Multilingual and code-switching ASR challenges for low resource Indian languages. In Proc. Interspeech, pages 2446–2450. find
  10. Injy Hamed, Nizar Habash, Slim Abdennadher, and Ngoc Thang Vu. 2022. ArzEn-ST: A three-way speech translation corpus for code-switched Egyptian Arabic-English. In Proc. WANLP, pages 119–130. find
  11. Amir Hussein, Dorsa Zeinali, Ondřej Klejch, Matthew Wiesner, Brian Yan, Shammur Chowdhury, Ahmed Ali, Shinji Watanabe, and Sanjeev Khudanpur. 2024. Speech collage: code-switched audio generation by collaging monolingual corpora. In Proc. ICASSP, pages 12006– 12010. find
  12. Fei Jia, Nithin Rao Koluguri, Jagadeesh Balam, and Boris Ginsburg. 2023. A compact end-to-end model with local and global context for spoken language identification. In Proc. Interspeech, pages 5321–5325. find
  13. Jee-weon Jung, Wangyou Zhang, Jiatong Shi, Zakaria Aldeneh, Takuya Higuchi, Barry-John Theobald, Ahmed Hussen Abdelaziz, and Shinji Watanabe. 2024. ESPnet-SPK: Full pipeline speaker embedding toolkit with reproducible recipes, self-supervised front-ends, and off-the-shelf models. arXiv preprint arXi… — [arXiv:2401.17230](https://arxiv.org/abs/2401.17230) find
  14. Jacob Kahn, Morgane Riviere, Weiyi Zheng, Evgeny Kharitonov, Qiantong Xu, Pierre-Emmanuel Mazaré, Julien Karadayi, Vitaliy Liptchinsky, Ronan Collobert, Christian Fuegen, et al. 2020. Libri-Light: A benchmark for ASR with limited or no supervision. In Proc. ICASSP, pages 7669–7673. find
  15. Shuyue Stella Li, Cihan Xiao, Tianjian Li, and Bismarck Odoom. 2023a. Simple yet effective code-switching language identification with multitask pre-training and transfer learning. arXiv preprint arXiv:2305.19759. — [arXiv:2305.19759](https://arxiv.org/abs/2305.19759) find
  16. Xinjian Li, Shinnosuke Takamichi, Takaaki Saeki, William Chen, Sayaka Shiota, and Shinji Watanabe. 2023b. YODAS: YouTube-oriented dataset for audio and speech. In Proc. IEEE ASRU, pages 1–8. find
  17. Dau-Cheng Lyu, Tien Ping Tan, Engsiong Chng, and Haizhou Li. 2010. SEAME: a mandarin-english code-switching speech corpus in southeast asia. In Proc. Interspeech, pages 1986– 1989. find
  18. C. Myers-Scotton. 1993. Duelling Languages: Grammatical Structure in Codeswitching. Clarendon Press. find
  19. Puyuan Peng, Brian Yan, Shinji Watanabe, and David Harwath. 2023. Prompting the hidden talent of web-scale speech models for zero-shot task generalization. In Proc. Interspeech, pages 396–400. find
  20. Yifan Peng, Shakeel Muhammad, Yui Sudo, William Chen, Jinchuan Tian, Chyi-Jiunn Lin, and Shinji Watanabe. 2025. OWSM v4: Improving open Whisper-style speech models via data scaling and cleaning. Proc. Interspeech, pages 2225– 2229. find
  21. Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, et al. 2024. Scaling speech technology to 1,000+ languages. Journal of Machine Learning Research, pages 1–52. find our card
  22. Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. MLS: A large-scale multilingual dataset for speech research. In Proc. Interspeech, pages 2757–2761. find our card
  23. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In Proc. ICML, pages 28492– 28518. find our card
  24. Pradeep Rangan, Sundeep Teki, and Hemant Misra. 2020. Exploiting spectral augmentation for code-switched spoken language identification. arXiv preprint arXiv:2010.07130. — [arXiv:2010.07130](https://arxiv.org/abs/2010.07130) find
  25. Jörgen Valk and Tanel Alumäe. 2021. VoxLingua107: a dataset for spoken language recognition. In Proc. IEEE SLT, pages 652–658. find
  26. Ewald van der Westhuizen and Thomas Niesler. 2018. A first south African corpus of multilingual code-switched soap opera speech. In Proc. LREC. find
  27. Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. 2021. VoxPopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. In Proc. ACL-IJCNLP. find our card
  28. Qingzheng Wang, Hye-jin Shim, Jiancheng Sun, and Shinji Watanabe. 2025. Geolocation-aware robust spoken language identification. In Proc. IEEE ASRU. find
  29. Brian Yan, Injy Hamed, Shuichiro Shimizu, Vasista Sai Lodagala, William Chen, Olga Iakovenko, Bashar Talafha, Amir Hussein, Alexander Polok, Kalvin Chang, et al. 2025. CS-FLEURS: A massively multilingual and code-switched speech dataset. In Proc. Interspeech, pages 743–747. find
  30. Miao Zhao, Yufeng Ma, Min Liu, and Minqiang Xu. 2021. The SpeakIn system for VoxCeleb speaker recognition challenge 2021. arXiv preprint arXiv:2109.01989. — [arXiv:2109.01989](https://arxiv.org/abs/2109.01989) find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yan-2026-cs-yodas-mined-dataset-wild/