Elamur Research⌕
☰
home / papers / 2026 / The Added Value of Metadata and Annotations: Evidence from Two Large-Scale, Naturalistic Corpus Studies

The Added Value of Metadata and Annotations: Evidence from Two Large-Scale, Naturalistic Corpus Studies

Summary

Speech recordings from media archives are plentiful and natural, but phoneticians often avoid them. Their audio quality varies, their annotations are sparse or inconsistent, and they rarely include reliable speaker metadata. The authors argue these corpora are still worth using, given suitable preparation. They show this through two case studies built on broadcast data originally gathered to train ASR systems. The first measures acoustic duration reduction of the hiatus /ia/ in European Portuguese as a function of stress and word position. The second is a sociophonetic study of word-final post-obstruent /r/ deletion in French, based only on ASR-derived pronunciation variants.

The paper focuses on the processing pipeline rather than the linguistic findings, which appear in companion papers. The pipeline covers removing noisy audio, checking segment boundaries, normalizing durations, resolving speaker identities and enriching metadata. The authors assess each step by how much it improves the fit (explained variance) of regression models. They conclude that scaling up phonetics depends on combining automatic and manual metadata augmentation with careful use of existing resources.

Key points

  • Data: The Portuguese study uses 114 h of TV and radio news and interviews. The French study combines 100 h of ESTER with 50 h of ETAPE.
  • French annotation quality:
    • Earlier work found a Cohen's Kappa of 0.832 between automatic and manual segmentation.
    • A manual check of 300 tokens labelled as /r/-deleted confirmed 98% of them.
  • Portuguese cleaning:
    • Files were flagged for background noise (1406), music (1138) or poor quality (565).
    • This reduced 9941 /ia/ tokens to a clean set of 6785 tokens in 1170 words.
    • After manual correction, boundaries were within 15 ms of the automatic ones for 79% of tokens.
  • Normalization: Durations were divided by articulation rate (phonemes per second) and then log-transformed. This corrected for the 10 ms frame discreteness and the skew of the raw values.
  • Speaker identity, two strategies:
    • Portuguese: auditory relabelling replaced 201 generic IDs with 3,658 new ones (2,422 in the clean set).
    • French: the data were restricted to 391 verifiable public figures, cutting about 26,000 /r/ tokens from 1,802 IDs down to roughly 14,000 tokens.

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 10.63317/4c4triganae3

@inproceedings{popescu2026added,
  title = {{The Added Value of Metadata and Annotations: Evidence from Two Large-Scale, Naturalistic Corpus Studies}},
  author = {Popescu, Anisia and Cronenberg, Johanna and Vasilescu, Ioana and Chitoran, Ioana and Lamel, Lori and Adda-Decker, Martine},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/4c4triganae3},
  url = {https://doi.org/10.63317/4c4triganae3}
}

References

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

  1. Martine Adda-Decker and Lori Lamel. 1999. Pronunciation Variants across System Configuration, Language and Speaking Style. Speech Communication, 29:83–98. find
  2. Emily Ahn and Eleanor Chodroff. 2022. VoxCommunis: A Corpus for Cross-linguistic Phonetic Analysis. In Proceedings of the 13th Conference on Language Resources and Evaluation, pages 5286–5294. find
  3. Alan W. Black. 2019. CMU Wilderness Multilingual Speech Dataset. In Proceedings of the 2019 International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5971– 5975. find
  4. Alebachew Chiche and Betselot Yitagesu. 2022. Part of speech tagging: a systematic review of deep learning and machine learning approaches. Journal of Big Data, 9(10). find
  5. Eleanor Chodroff and Miao Zhang. preprint. A Crosslinguistic Corpus Phonetic Analysis of Intrinsic Vowel Duration. arXiv. find
  6. Jacob Cohen. 1960. A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement, 20(1):37–46. find
  7. Abigail C. Cohn and Margaret E. L. Renwick. 2019. Doing Phonology in the Age of Big Data. In Cornell Working Papers in Phonetics and Phonology 2019, pages 1–36. find
  8. John Coleman, Mark Liberman, Greg Kochanski, Lou Burnard, and Jiahong Yuan. 2011. Mining a year of speech. In Proceedings of New Tools and Methods for Very-Large-Scale Phonetics Research, pages 16–19. find
  9. BNC Consortium. 2007. The British National Corpus, XML Edition. find
  10. Johanna Cronenberg, Ioana Chitoran, Lori Lamel, and Ioana Vasilescu. 2024. Crosslinguistic Comparison of Acoustic Variation in the Vowel Sequences /ia/ and /io/ in Four Romance Languages. In Proceedings of the 25th Interspeech Conference, pages 3689–3693. find
  11. Johanna Cronenberg, Lori Lamel, and Ioana Chitoran. In prep. Reduction of the Hiatus /ia/ in Naturalistic European Portuguese. find
  12. Thiago Fraga-Silva, Jean-Luc Gauvain, and Lori Lamel. 2011. Lattice-based Unsupervised Acoustic Model Training. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP). find
  13. Thiago Fraga-Silva, Jean-Luc Gauvain, and Lori Lamel. 2013. Interpolation of acoustic models for speech recognition. In Proceedings of the Interspeech, pages 3347–3351. find
  14. Sylvain Galliano, Edouard Geoffrois, Guillaume Gravier, Jean-Francois Bonastre, Djamel Mostefa, and Khalid Choukri. 2006. Corpus description of the ESTER evaluation campaign for the rich transcription of French broadcast news. In Proc. find
  15. LREC 2006, volume 6, pages 315–320. find
  16. Guilherme D. Garcia. 2017. Weight Gradience and Stress in Portuguese. Phonology, 34(1):41–79. find
  17. John J. Godfrey and Edward Holliman. 1993. Switchboard-1 Release 2 LDC97S62. Web Download. find
  18. Guillaume Gravier, Gilles Adda, Niklas Paulson, Mathieu Carré, Aude Giraudel, and Olivier Galibert. 2012. The ETAPE corpus for the evaluation of speech-based TV content processing in the French language. In Proc. find
  19. Simone Harmath-de Lemos. 2022. Detecting word-level stress in continuous speech: A case study of brazilian portuguese. Journal of Portuguese Linguistics, 20(101). find
  20. Jonathan Harrington, Sallyanne Palethorpe, and Catherine I. Watson. 2000. Does the Queen Nature, speak the Queen’s English? 408(6815):927–928. find
  21. Jennifer B. Hay, Stefanie Jannedy, and Norma Mendoza-Denton. 1999. Oprah and /ay/: Lexical Frequency, Referee Design and Style. In Proceedings of the 14th International Congress of Phonetic Sciences, pages 1389–1392. find
  22. Alexandra Kuznetsova, Per B. Brockhoff, and Rune H. B. Christensen. 2017. lmerTest package: Tests in Linear Mixed Effects Models. Journal of Statistical Software, 82(13):1–26. find
  23. William Labov. 1972. Blackwell. find
  24. Lori Lamel and Jean-Luc Gauvain. 1992. Continuous Speech Recognition at LIMSI. In Proceedings of the Final Review of the DARPA ANNT Speech Program, pages 1–7. find
  25. Mark Y. Liberman. 2019. Corpus Phonetics. Annual Review of Linguistics, 5(1):91–107. find
  26. Shuo Liu, Gil Keren, Emilia Parada-Cabaleiro, and Björn Schuller. 2021. N-HANS: A Neural Network-Based Toolkit for in-the-wild Audio Enhancement. Multimedia Tools and Applications, 80(18):28365–28389. find
  27. Brian MacWhinney. 2000. The CHILDES Project: Tools for Analyzing Talk, third edition edition. Lawrence Erlbaum Associates, Mahwah, NJ. find
  28. Anisia Popescu, Mathilde Hutin, Lori Lamel, Ioana Vasilescu, and Martine Adda-Decker. 2023. Stop devoicing and place of articulation: a cross-linguistic study using large-scale corpora. In Proceedings of the 20th ICPhS, pages 3186 – 3190. find
  29. Anisia Popescu, Lori Lamel, Marc Evrard, and Ioana Vasilescu. 2025. Tracking /r/ deletion: Forced alignment of pronunciation variants and sociophonetic insights into post-obstruent final /r/ in french. In Proceedings of the Interspeech, pages 2945 – 2949. find
  30. Rasmus Puggaard-Rode, Camilla Søballe Horslund, and Henrik Jørgensen. 2022. The Rarity of Intervocalic Voicing of Stops in Danish Spontaneous Speech. Laboratory Phonology, 13(1):1–47. find
  31. Hugo Quené. 2013. Longitudinal Trends in Speech Tempo: The case of Queen Beatrix. The Journal of the Acoustical Society of America, 133(6):EL452–EL457. find
  32. Morgan Sonderegger, Jane Stuart-Smith, Jeff Mielke, and The SPADE Consortium. 2023. How variable are English sibilants? In Proceedings of the 20th International Congress of Phonetic Sciences, pages 3196–3200. find
  33. Joseph A. Stanley, Lisa Morgan Johnson, and Earl Kjar Brown. 2025. Testing the Effect of Speech Separation on Vowel Formant Estimates. Linguistics Vanguard. find
  34. Martin A. Stoffel, Shinichi Nakagawa, and Holger Schielzeth. 2021. partR2: Partitioning R2 in Generalized Linear Mixed Models. PeerJ. find
  35. Ioana Vasilescu, Yaru Wu, Adèle Jatteau, Martine Adda-Decker, and Lori Lamel. 2020. Alternances de Voisement et Processus de Lénition et de Fortition: Une Étude Automatisée de Grands Corpus en Cinq Langues Romanes. Traitement Automatique des Langues, 61(1):11–36. find
  36. Yaru Wu. 2018. Étude de la réduction segmentale en français parlé à travers différents styles : apports des grands corpus et du traitement automatique de la parole à l’étude du schwa, du /K/ et des réductions à segments multiples. Phd thesis, Université Sorbonne Nouvelle - Paris 3. find
  37. Yaru Wu, Lihu Chen, Benjamin Elie, Fabian Suchanek, Ioana Vasilescu, and Lori Lamel. 2023. Who’s speaking? predicting speaker profession from speech. In Proc. find
  38. Yaru Wu, Fabian Suchanek, Ioana Vasilescu, Lori Lamel, and Martine Adda-Decker. 2022. Using a knowledge base to automatically annotate speech corpora and to identify sociolinguistic variation. In Proc. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/popescu-2026-added-value-metadata-annotations-evidence/