Elamur Research⌕
☰
home / papers / 2026 / Simultaneous Speech-to-Text Translation Web Application for Estonian

Simultaneous Speech-to-Text Translation Web Application for Estonian

Paper · 2026 · Bohdan Podziubanchuk, Aivo Olev, Jiaming Kong et al.

Summary

The paper presents an open-source web application for live speech-to-text translation centred on Estonian. Estonian is a small Uralic language, and growing use of English in fields like higher education raises fears of domain loss. The system translates spoken Estonian into English, Russian, and Ukrainian text, and English speech into Estonian. It uses a cascade. First, a streaming Zipformer transducer ASR model produces punctuated, capitalized Estonian transcripts. Then a fine-tuned LLM translates them in five-word chunks, framing the task as a multi-turn dialogue (following Wang et al., 2025). Each source chunk is a user turn and each partial translation an assistant turn. This lets the model reuse its key–value cache and run on standard LLM inference tools. Training data for this chunked behaviour was built from word-aligned parallel sentences, so no target word appears before its aligned source word. The system reaches 10.2% WER and 26.1 BLEU for Estonian→English with 3–6 seconds of latency, clearly beating SeamlessStreaming. The released code and data-generation scripts make it adaptable to other language pairs, including Ukrainian-related ones.

Key points

  • ASR: a roughly 150M-parameter Zipformer model trained with Icefall on three sources:
    • about 1334 h from the TalTech Estonian Speech Dataset 1.0;
    • about 4000 h of ETV broadcasts auto-transcribed with fine-tuned Whisper large-v3-turbo;
    • a 500 h re-transcribed Gigaspeech subset to handle English terms mixed into Estonian speech.
  • MT: Hunyuan-7B-MT fully fine-tuned for one epoch on 500K sentence pairs sampled from the SynEst corpus, covering et↔en, et↔uk, and et↔ru. Word alignments came from XLM-R, and training took 61 h on eight AMD MI250X GPUs.
  • Results (BLEU / BLEURT):
    • Estonian→English: 26.1 / 0.56, versus 13.8 for Seamless Streaming and 31.9 for an offline Whisper + Neurotõlge cascade.
    • Estonian→Russian: 22.3.
    • English→Estonian: 11.1, held back by a weak English ASR model.
  • Base model choice: Hunyuan-7B-MT outperformed Llama-3.1-8B-EstLLM, suggesting a translation-specialised LLM beats a language-specific general one for this task.
  • Deployment: a Next.js/React frontend linked over WebSocket to a Rust backend (Axum, sherpa-onnx). Extras include sharing translations to mobile devices and streaming output to OBS Studio. The live demo runs at est2eng.cs.taltech.ee.

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

🎓Tallinn University of TechnologyTallinn · Estonia5 works · 62,036 citations

 Fields, topics and concepts

📚 Shelves 4The 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 works

🧭 Topics 1Research subjects and languages the work is about.Speech & spoken language · 480 works

Cite

Cite as Bohdan Podziubanchuk, Aivo Olev, Jiaming Kong and Tanel Alumäe. 2026. Simultaneous Speech-to-Text Translation Web Application for Estonian. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 3: System Demonstrations).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-demo.40

@inproceedings{podziubanchuk2026simultaneous,
  title = {{Simultaneous Speech-to-Text Translation Web Application for Estonian}},
  author = {Bohdan Podziubanchuk, Aivo Olev, Jiaming Kong and Tanel Alumäe},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-demo.40},
  url = {https://doi.org/10.18653/v1/2026.eacl-demo.40}
}

References

12 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Elizaveta Korotkova, Taido Purason, Agnes Luhtaru, and Mark Fishel. 2024. Multilinguality or back-translation? A case study with Estonian. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 11838–11848,… find
  2. Tanel Alumäe, Joonas Kalda, Külliki Bode, and Martin Kaitsa. 2023. Automatic closed captioning for Estonian live broadcasts. In Proceedings of the 24th Nordic Conference on Computational Linguis-Evgeny Matusov, Gregor Leusch, Oliver Bender, and Hermann Ney. 2005. Evaluating machine translation outpu… find
  3. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318.
  4. Mao Zheng, Zheng Li, Bingxin Qu, Mingyang Song, Yang Du, Mingrui Sun, and Di Wang. 2025. Hunyuan-MT technical report. Preprint, arXiv:2509.05209.
  5. Amy Pu, Hyung Won Chung, Ankur Parikh, Sebastian Gehrmann, and Thibault Sellam. 2021. Learning compact metrics for MT. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 751–762, Online and Punta Cana, Dominican Republic. Association for Computational Li…
  6. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust speech recognition via large-scale weak supervision. Preprint, arXiv:2212.04356.
  7. Seamless Communication, Loïc Barrault, Yu-An Chung, Mariano Coria Meglioli, David Dale, Ning Dong, Mark Duppenthaler, Paul-Ambroise Duquenne, Brian Ellis, Hady Elsahar, Justin Haaheim, John Hoffman, Min-Jae Hwang, Hirofumi Inaguma, Christopher Klaiber, Ilia Kulikov, Pengwei Li, Daniel Licht, Jean Ma… — Seamless: Multilingual Expressive and Streaming Speech Translation
  8. Thibault Sellam, Dipanjan Das, and Ankur Parikh. 2020. BLEURT: Learning robust metrics for text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7881–7892, Online. Association for Computational Linguistics.
  9. Tiia Sildam, Andra Velve, and Tanel Alumäe. 2024. Finetuning end-to-end models for Estonian conversational spoken language translation. In Proceedings of the Seventh Workshop on Technologies for Machine Translation of Low-Resource Languages (LoResMT 2024), pages 166–174, Bangkok, Thailand. Associati… find
  10. Andre Tättar, Taido Purason, Hele-Andra Kuulmets, Agnes Luhtaru, Liisa Rätsep, Maali Tars, Marcis Pinnis, Toms Bergmanis, and Mark Fishel. 2022. Open and competitive multilingual neural machine translation in production. Proceedings of Baltic HLT. find
  11. Minghan Wang, Thuy-Trang Vu, Yuxia Wang, Ehsan Shareghi, and Gholamreza Haffari. 2025. Conversational SimulMT: Efficient simultaneous translation with large language models. In Proceedings of the 22nd International Conference on Spoken Language Translation (IWSLT 2025), pages 93–105, Vienna, Austria… find
  12. Zengwei Yao, Liyong Guo, Xiaoyu Yang, Wei Kang, Fangjun Kuang, Yifan Yang, Zengrui Jin, Long Lin, and Daniel Povey. 2024. Zipformer: A faster and better encoder for automatic speech recognition. Preprint, arXiv:2310.11230. — [arXiv:2310.11230](https://arxiv.org/abs/2310.11230) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/podziubanchuk-2026-simultaneous-speech-text-translation-web/