Elamur Research⌕
☰
home / papers / 2026 / Exploring Cross-Lingual Voice Conversion Methods for Anonymizing Low-Resource Text-to-Speech

Exploring Cross-Lingual Voice Conversion Methods for Anonymizing Low-Resource Text-to-Speech

Paper · 2026 · Shenran Wang, Aidan Pine and Mengzhe Geng

Summary

Neural TTS for small language communities raises a privacy concern: the few people who record training data are often well known, and a model that can say anything in their voice could harm their reputation or lend credibility to scams. The authors test ways to hide the original speakers' identity while keeping synthetic speech natural. They work with StyleTTS2 and two Canadian Indigenous languages, nêhiyawêwin (Plains Cree) and SENĆOTEN. They compare three families of methods:

  • editing style embeddings at inference time,
  • post-processing output with off-the-shelf SeedVC models,
  • training multilingual models with English LibriTTS-R so that an English reference speaker can supply the voice.

All evaluation is automatic, since human evaluation was not feasible. Multilingual training with added language embeddings gave the most reliable balance across both languages: it lowered speaker similarity and kept quality high. SeedVC degraded quality, and models without language embeddings produced a strong English accent. The study gives practical guidance for ethically deploying TTS in low-resource communities. It also shows that cross-lingual voice conversion needs better evaluation metrics.

Key points

  • Data: nêhiyawêwin has about 8.2 h of training audio from 8 speakers, and SENĆOTEN about 3.9 h from 2 speakers. Each was mixed with about 53 h of LibriTTS-R train-clean-100.
  • Metrics: PESQ and SI-SDR were predicted with torchaudio-squim, and MOS with UTMOS. Speaker similarity (SECS, lower is better) came from Resemblyzer. MMS-LID confidence served as a nêhiyawêwin accent proxy (CRK-Conf).
  • nêhiyawêwin results: The CRK-ENG model with language embeddings reached MOS 4.03, SECS 0.63 and CRK-Conf 0.67. The no-conversion baseline had SECS 0.87. Without language embeddings, SECS fell to 0.52, but CRK-Conf collapsed to 0.003, indicating an English accent.
  • SeedVC and SENĆOTEN: SeedVC variants (v1-base/small/xlsr, v2) cut SECS to roughly 0.52–0.58 but often hurt quality, with v1-base dropping to MOS 2.34. For SENĆOTEN, the inference-time "Custom" setting only reached SECS 0.74.
  • Model scope and release: Adding more languages did not help, as CRK-ENG beat CRK-STR-ENG on most metrics. Model weights stay with the communities, while the code is public on GitHub (EveryVoiceTTS).

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

 Institutions

🏛️National Research Council CanadaOttawa · Canada59 works · 21,677 citations🎓University of British ColumbiaVancouver · Canada98 works · 34,805 citations

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSignal Processing · 452 worksMedicine · 241 worksPhysiology · 29 works

🧭 Topics 3Research subjects and languages the work is about.Multilingual & cross-lingual NLP · 828 worksSpeech & spoken language · 480 worksLow-resource languages · 267 works

Cite

Cite as Shenran Wang, Aidan Pine and Mengzhe Geng. 2026. Exploring Cross-Lingual Voice Conversion Methods for Anonymizing Low-Resource Text-to-Speech. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-short.16

@inproceedings{wang2026exploring,
  title = {{Exploring Cross-Lingual Voice Conversion Methods for Anonymizing Low-Resource Text-to-Speech}},
  author = {Shenran Wang, Aidan Pine and Mengzhe Geng},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-short.16},
  url = {https://doi.org/10.18653/v1/2026.eacl-short.16}
}

References

7 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Chung-Ming Chien, Jheng-Hao Lin, Chien-yu Huang, Po-chun Hsu, and Hung-yi Lee. 2021. Investigating on incorporating pretrained and learnable speaker representations for multi-speaker multi-style text-to-speech. In ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Proce… find
  2. Wiebke Hutiri, Orestis Papakyriakopoulos, and Alice Xiang. 2024. Not my voice! a taxonomy of ethical and safety harms of speech generators. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’24, page 359–376, New York, NY, USA. Association for Computing M… find
  3. Aidan Pine, Erica Cooper, David Guzmán, Eric Joanis, Anna Kazantseva, Ross Krekoski, Roland Kuhn, Samuel Larkin, Patrick Littell, Delaney Lothian, Ak-wiratékha’ Martin, Korin Richmond, Marc Tessier, Cassia Valentini-Botinhao, Dan Wells, and Junichi Yamagishi. 2025. Speech generation for indigenous l… find
  4. Aidan Pine, Dan Wells, Nathan Brinklow, Patrick Littell, and Korin Richmond. 2022. Requirements and motivations of low-resource speech synthesis for language revitalization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7346… find
  5. Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, and Michael Auli. 2024. Scaling Speech Technology to 1,000+ Languages. Journal of M…
  6. Takaaki Saeki, Detai Xin, Wataru Nakata, Tomoki Koriyama, Shinnosuke Takamichi, and Hiroshi Saruwatari. 2022. Utmos: Utokyo-sarulab system for voicemos challenge 2022. In Interspeech 2022, pages 4521–4525. find
  7. Jonathan Shen, Ruoming Pang, Ron J. Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, R. J. Skerry-Ryan, Rif A. Saurous, Yannis Agiomyrgiannakis, and Yonghui Wu. 2018. Natural TTS synthesis by conditioning WaveNet on Mel spectrogram predictions. pages 4779–478… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-exploring-cross-lingual-voice-conversion/