Elamur Research⌕
☰
home / papers / 2026 / How Much Data for Stable Formant Values? Pipeline for Convergence Detection Based on Read Speech

How Much Data for Stable Formant Values? Pipeline for Convergence Detection Based on Read Speech

Summary

Vowel formant studies usually measure short, isolated utterances recorded in a lab. This leaves an open question: how many vowel tokens are needed before a speaker's average F1, F2 and F3 stop changing in a meaningful way? The authors address it with a large corpus of English audiobooks, 3,384 chapters totalling roughly 942 hours. The material comes from six solo LibriVox narrators plus a King James Bible recording. They built a fully automated, reproducible pipeline with four stages:

  • ASR transcription with faster-whisper
  • phoneme alignment with the Montreal Forced Aligner
  • formant extraction at vowel midpoints with Praat via Parselmouth
  • IQR-based outlier filtering

They tracked cumulative formant means token by token and compared six methods for detecting the point where the mean levels off. Across vowels, means typically settled after about 60–230 tokens per phoneme, depending on vowel quality and speaker gender. CUSUM gave the most consistent and informative stabilization estimates. The work offers practical sample-size guidance for formant research on non-laboratory speech. It also provides a baseline pipeline that could be applied to other speech types.

Key points

  • Data: Six LibriVox narrators were selected for catalog size and for gender and dialect coverage. The King James Bible recording, read by Winfred Henson and taken from eBible.org, has 1,189 chapters of about 650 words each. Audio was converted with pydub to 16 kHz mono 16-bit WAV.
  • Tools:
    • Transcription used faster-whisper v1.1.0 with the small model and word timestamps.
    • Alignment used MFA 3.2.1 with english_mfa models. MFA was chosen over WebMAUS because it had lower MAE on the TIMIT TEST subset.
    • Formants came from the Burg algorithm, with a ceiling of 5,000 Hz for male and 5,500 Hz for female speakers.
  • Preprocessing: IQR filtering within each chapter removed about 4% of vowel tokens as outliers.
  • Methods compared: FTW, CUSUM, Two-Sample t-Test, CI Shrinkage, PWLF and BinSeg, each tested across several parameter settings.
  • Results: Stabilization occurred at roughly 60–230 tokens per vowel phoneme, or around 7 minutes of speech within a single recording. CUSUM was judged the most reliable detector.

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 10.63317/3mgdob6szndp

@inproceedings{sward2026how,
  title = {{How Much Data for Stable Formant Values? Pipeline for Convergence Detection Based on Read Speech}},
  author = {Sward, Kayla and Sjons, Johan and Ekstrom, Axel G.},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/3mgdob6szndp},
  url = {https://doi.org/10.63317/3mgdob6szndp}
}

References

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

  1. Emily P Ahn, Gina-Anne Levow, Richard A Wright, and Eleanor Chodroff. 2023. An outlier analysis of vowel formants from a corpus phonetics pipeline. find
  2. INTERSPEECH 2023, pages 2573– 2577. find
  3. Paul Boersma. 2001. Praat, a system for doing phonetics by computer. Glot. Int., 5(9):341–345. find
  4. Paul Boersma and David Weenink. 2020. Praat: Sound: To formant (burg)... https://www.fon.hum.uva.nl/praat/manual/Sound__To_Formant__burg____.html. ↗
  5. Mert Can Cakmak and Nitin Agarwal. 2024. High-speed transcript collection on multimedia platforms: Advancing social media research through parallel processing. In 2024 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), pages 857–860. IEEE. eBible.org. World english… find
  6. Carlos Fortea-Sanchis and Javier Escrig-Sos. 2019. Quality control techniques in surgery: application of cumulative sum (cusum) charts. Cirugía Española (English Edition), 97(2):65–70. find
  7. John S Garofolo, Lori F Lamel, William M Fisher, David S Pallett, Nancy L Dahlgren, Victor Zue, and Jonathan G Fiscus. 1993. Timit acoustic-phonetic continuous speech corpus. (No Title). find
  8. William J Hardcastle, John Laver, and Fiona E Gibbon. 2012. The handbook of phonetic sciences. John Wiley & Sons. find
  9. John D Healy. 1987. A note on multivariate cusum procedures. Technometrics, 29(4):409–412. find
  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. find
  11. Yannick Jadoul, Bill Thompson, and Bart De Boer. 2018. Introducing parselmouth: A python interface to praat. Journal of Phonetics, 71:1–15. find
  12. Charles F Jekel and Gerhard Venter. 2019. pwlf: A python library for fitting 1d continuous piecewise linear functions. URL: https://github.com/cjekel/piecewise_linear_fit_py. ↗
  13. Jonathan Jibson. 2019. Variability of formant values at different time points of vowels. In Proceedings of Meetings on Acoustics, volume 39. AIP Publishing. find
  14. Guillaume Klein. 2024. faster-whisper: Audio preprocessing source code. https://github.com/SYSTRAN/faster-whisper/blob/ master/faster_whisper/audio.py. ↗
  15. Sang Gyu Kwak and Jong Hae Kim. 2017. Central limit theorem: the cornerstone of modern statistics. Korean journal of anesthesiology, 70(2):144. find
  16. William Labov, Sharon Ash, and Charles Boberg. 2006. The atlas of North American English: Phonetics, phonology and sound change. Mouton de Gruyter. find
  17. LibriVox. Librivox free public domain audiobooks. Accessed March 2025. find
  18. LibriVox Community. 2022. Newbie guide to recording. find
  19. Daniel E Lieberman, Robert C McCarthy, Karen M Hiiemae, and Jeffrey B Palmer. 2001. Ontogeny of postnatal hyoid and larynx descent in humans. Archives of oral biology, 46(2):117–128. find
  20. Dominik Macháček, Raj Dabre, and Ondřej Bojar. 2023. Turning whisper into real-time transcription system. arXiv preprint arXiv:2307.14743. find
  21. Michael McAuliffe, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger. 2017. Montreal forced aligner: Trainable text-speech alignment using kaldi. In Interspeech, volume 2017, pages 498–502. find
  22. Michael McAuliffe and Morgan Sonderegger. 2024. Montreal forced aligner: Corpus formats and structure. https://montreal-forced-aligner.readthedocs.io/en/v3.2.1/user_ guide/corpus_structure.html. ↗
  23. Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206–5210. IEEE. find
  24. 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. find
  25. Florian Schiel. 1999. Automatic phonetic transcription of non-prompted speech. find
  26. Lee Schruben. 1983. Confidence interval estimation using standardized time series. Operations Research, 31(6):1090–1108. find
  27. Xuesheng Shi, Colin Gallagher, Robert Lund, and Rebecca Killick. 2022. A comparison of single and multiple changepoint techniques for time series data. Computational Statistics & Data Analysis, 170:107433. find
  28. Charles Truong, Laurent Oudre, and Nicolas Vayatis. 2020. Selective review of offline change point detection methods. Signal Processing, 167:107299. find
  29. Xiaolan L Wang, Qiuzi H Wen, and Yuehua Wu. 2007. Penalized maximal t test for detecting undocumented mean change in climate data series. Journal of Applied Meteorology and Climatology, 46(6):916–931. find
  30. David Weenink. 2015. Improved formant frequency measurements of short segments. In Proceedings of the 18th International Congress of Phonetic Sciences (ICPhS), Glasgow, UK. The International Phonetic Association. find
  31. Rolf Werner, Dimitar Valev, Dimitar Danov, and Veneta Guineva. 2015. Study of structural break points in global and hemispheric temperature series by piecewise regression. Advances in Space Research, 56(11):2323–2334. find
  32. Hongchen Wu, Jiwon Yun, Xiang Li, Huiyi Huang, and Chuandong Liu. 2023. Using a forced aligner for prosody research. Humanities and Social Sciences Communications, 10(1):1–13. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sward-2026-how-much-data-stable-formant/