Elamur Research⌕
☰
home / papers / 2026 / TTSVowelViz: A Tool for Visualising Text-to-Speech Model Training via Vowel Spaces

TTSVowelViz: A Tool for Visualising Text-to-Speech Model Training via Vowel Spaces

Summary

A falling loss curve tells TTS developers when a model has converged, but not which speech sounds it has actually learned. TTSVowelViz fills this gap by plotting vowel spaces while a model is still training, bringing phonetic knowledge into model development. At user-chosen training steps, the tool synthesises evaluation word lists with the intermediate model and aligns phonemes using WebMAUS. It then measures F1 and F2 with Parselmouth (Praat) and draws interactive Plotly charts on the Bark scale. Static plots show one averaged formant pair per vowel. Dynamic plots trace formant trajectories over the vowel's duration, which suits diphthongs. Because the tool needs only the model's synthesis routine, it works with any framework. To validate it, the authors took a Tacotron 2 voice trained on General American English, fine-tuned it to speak New Zealand English, and ran a listening test. Formant movements of specific vowels in the plots matched what listeners heard, so the tool made the accent shift visible. TTSVowelViz does not replace perception tests. It adds a real-time, interpretable check that can guide training settings and dataset refinement.

Key points

  • The tool is released as the Python package ttsvowelviz on PyPI, with an interactive web demo. Unlike earlier proof-of-concept studies that analysed saved checkpoints after training, it visualises vowel spaces during training.
  • The pipeline runs synthesis, then WebMAUS segmentation into TextGrids, then Parselmouth formant extraction, then Plotly plots. Users can extract formants at any number of time points, and averaging over all occurrences of a vowel reduces segmentation and extraction noise.
  • In the experiment, Tacotron 2 was pre-trained for 120000 steps on LJSpeech (13100 clips, ≈24 h, female GAE speaker). It was then fine-tuned on the Mansfield corpus (831 clips, ≈1 h, female NZE speaker) for 28000 steps, taking about 36 h on a Tesla K80 at learning rate 0.001.
  • Vowel spaces were plotted at 8 training steps (0 to 28000), using FLEECE, THOUGHT and START as point vowels. Monophthongs were measured at their midpoint and diphthongs at 20%, 50% and 80% of their duration. Twenty-five word lists were built for each vowel class, labelled with Wells' lexical sets.
  • The formant shifts shown in the plots agreed with the perception test responses, which supports vowel-space plots as an interpretability aid for TTS training.

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 10.63317/57peripccxqx

@inproceedings{udawatta2026ttsvowelviz,
  title = {{TTSVowelViz: A Tool for Visualising Text-to-Speech Model Training via Vowel Spaces}},
  author = {Udawatta, Pasindu and James, Jesin and T, Balamurali B and Watson, Catherine Inez and Nicholas, Ake and Abeysinghe, Binu Nisal},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/57peripccxqx},
  url = {https://doi.org/10.63317/57peripccxqx}
}

References

35 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Binu Abeysinghe, Jesin James, Catherine I. Watson, and Felix Marattukalam. 2022. Visualising Model Training via Vowel Space for Text-To-Speech Systems. In Interspeech, pages 511– 515. find
  2. Sven Albrecht, Rewa Tamboli, Stefan Taubert, Maximilian Eibl, Günter D. Rey, and Josef Schmied. 2022. Towards a Vowel Formant Based Quality Metric for Text-to-Speech Systems: Measuring Monophthong Naturalness. In IEEE International Conference on Computational Intelligence and Virtual Environments for Measurement Systems and Applications. find
  3. Henry An, Jesin James, Catherine Watson, and Binu Abeysinghe. 2024. Visualising Vowel Dynamics for the Training of Text to Speech Systems. In Australasian International Conference on Speech Science and Technology, pages 197– 201. find
  4. David Brett. 2004. Computer generated feedback on vowel production by learners of English as a second language. ReCALL, 16(1):103–113. ↗
  5. Michael Carey. 2004. CALL Visual Feedback for Pronunciation of Vowels. CALICO Journal, 21(3):571–601. ↗
  6. Xinyuan Chao, Charbel El-Khaissi, Nicholas Kuo, Priscilla K. John, and Hanna Suominen. 2020. An Automatic Vowel Space Generator for Language Learner Pronunciation Acquisition and Correction. In Workshop of the Australasian Language Technology Association, pages 54–64. find
  7. Felicity Cox, Joshua Penney, and Sallyanne Palethorpe. 2024. Australian English Monophthong Change across 50 Years: Static versus Dynamic Measures. Languages, 9(3). ↗
  8. Anushiya R. Gladston and K. V. Pradeep. 2023. Exploring Solutions for Text-to-Speech Synthesis of Low-Resource Languages. In International Conference on Signal Processing and Communication, pages 168–172. find
  9. Jonathan Harrington, Felicity Cox, and Zoe Evans. 1997. An acoustic phonetic study of broad, general, and cultivated Australian English vowels. Australian Journal of Linguistics, 17(2):155–184. ↗
  10. James Hillenbrand, Laura A. Getty, Michael J. Clark, and Kimberlee Wheeler. 1995. Acoustic characteristics of American English vowels. The Journal of the Acoustical Society of America, 97(5):3099–3111. ↗
  11. Ewa Jacewicz, Robert A. Fox, and Joseph Salmons. 2007. Vowel space areas across dialects and gender. In International Congress of Phonetic Sciences, pages 1465–1468. find
  12. Yannick Jadoul, Bill Thompson, and Bart de Boer. 2018. Introducing Parselmouth: A Python interface to Praat. Journal of Phonetics, 71. ↗
  13. Sandra Jansen. 2019. Change and stability in GOOSE, GOAT and FOOT: back vowel dynamics in Carlisle English. English Language and Linguistics, 23(1). find
  14. Allard Jongman, Marios Fourakis, and Joan A. Sereno. 1989. The Acoustic Vowel Space of Modern Greek and German. Language and Speech, 32(3):221–248. ↗
  15. Michael Kiefte and Terrance M. Nearey. 2017. Modeling consonant-context effects in a large database of spontaneous speech recordings. The Journal of the Acoustical Society of America, 142(1):434–443. ↗
  16. Yogesh Kumar, Apeksha Koul, and Chamkaur Singh. 2023. A deep learning approaches in text-to-speech system: a systematic review and recent research perspective. Multimedia Tools and Applications, 82:15171–15197. ↗
  17. Philipos C. Loizou. 2011. Speech Quality Assessment. In Multimedia Analysis, Processing and Communications, volume 346, pages 623–654. Springer Berlin Heidelberg. ↗
  18. Gordon E. Peterson and Harold L. Barney. 1952. Control Methods Used in a Study of the Vowels. The Journal of the Acoustical Society of America, 24(2):175–184. ↗
  19. Margaret Renwick and Joseph Stanley. 2020. Modeling dynamic trajectories of front vowels in the American South. The Journal of the Acoustical Society of America, 147:579–595. ↗
  20. Brooke Ross, Elaine Ballard, and Catherine Watson. 2023. Front vowel lowering in New Zealand English: investigating a regional sound change. In International Congress of Phonetic Sciences, pages 3666–3670. find
  21. Brooke C. Ross. 2024. A new look at sound change in New Zealand English. Ph.D. thesis, The University of Auckland. find
  22. Steven Sandoval, Visar Berisha, Rene L. Utianski, Julie M. Liss, and Andreas Spanias. 2013. Automatic assessment of vowel space area. The Journal of the Acoustical Society of America, 134(5):EL477–EL483. ↗
  23. Florian Schiel. 1999. Automatic Phonetic Transcription of Non-Prompted Speech. In International Congress of Phonetic Sciences, pages 607–610. find
  24. Jonathan Shen, Ruoming Pang, Ron J. Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, Rj Skerrv-Ryan, Rif A. Saurous, Yannis Agiomvrgiannakis, and Yonghui Wu. 2018. Natural TTS Synthesis by Conditioning Wavenet on MEL Spectrogram Predictions. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4779–4783. find
  25. Joseph A. Stanley, Margaret E. L. Renwick, Katherine I. Kuiper, and Rachel M. Olsen. 2021. Back Vowel Dynamics and Distinctions in Southern American English. Journal of English Linguistics, 49(4):389–418. ↗
  26. Pasindu Udawatta, Jesin James, Catherine Watson, Ake Nicholas, and B. T. Balamurali. 2025. Can Vowel Space Visualisation Reveal How Text-to-Speech Systems Learn? In Conference of the Linguistic Society of New Zealand, page 91. ↗
  27. Catherine Watson and Jonathan Harrington. 1999. Acoustic evidence for dynamic formant trajectories in Australian English vowels. The Journal of the Acoustical Society of America, 106:458–468. ↗
  28. Catherine Watson, Brooke Ross, Elaine Ballard, Helen Charters, Richard Arnold, and Miriam Meyerhoff. 2018. Preliminary Investigations into Sound Change in Auckland. In Australasian International Conference on Speech Science and Technology, pages 17–20. find
  29. Catherine I. Watson, Jonathan Harrington, and Zoe Evans. 1998. An acoustic comparison between New Zealand and Australian English vowels. Australian Journal of Linguistics, 18(2):185– 207. ↗
  30. Catherine I. Watson, Peter J. Keegan, Margaret A. Maclagan, Ray Harlow, and Jeanette King. 2017. The Motivation and Development of MPAi, a Māori Pronunciation Aid. In Interspeech, pages 2063–2067. find
  31. Catherine I. Watson, Margaret A. Maclagan, Jeanette King, Ray Harlow, and Peter J. Keegan. 2016. Sound change in Māori and the influence of New Zealand English. Journal of the International Phonetic Association, 46(2):185–218. ↗
  32. John C. Wells. 1982a. Accents of English 1: An Introduction, volume 1. Cambridge University Press. find
  33. John C. Wells. 1982b. Accents of English 3: Beyond the British Isles, volume 3. Cambridge University Press. 10. find
  34. Language Resource References Keith Ito and Linda Johnson. 2017. The LJ Speech Dataset. PID https://keithito.com/LJ-Speech-Dataset/. ↗
  35. Catherine Watson and Amélie Marchi. 2014. Resources Created For Building New Zealand English Voices. In Australasian International Conference on Speech Science and Technology, pages 92–95. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/udawatta-2026-ttsvowelviz-tool-visualising-text-speech/