Elamur Research⌕
☰
home / papers / 2026 / MD_NLP: Reconstructing an Australian English Heritage Dialect Corpus from the Mitchell-Delbridge Recordings through LLM-Assisted Speaker Attribution

MD_NLP: Reconstructing an Australian English Heritage Dialect Corpus from the Mitchell-Delbridge Recordings through LLM-Assisted Speaker Attribution

Paper · 2026 · Steven Coats

Summary

The Mitchell–Delbridge (MD) archive holds recordings of Australian secondary school pupils made in 1959/60. Its spontaneous narrative part had gone almost unused because there were no transcripts and the audio quality varies widely. Earlier studies drew only on the read sentences and word lists.

MD_NLP turns these narratives into a usable corpus through a layered pipeline:

  • WhisperX produces the transcript.
  • Pyannote separates speakers.
  • An LLM (Gemini) reasons over the dialogue to fix speaker labels and assign interviewer or pupil roles.
  • The Montreal Forced Aligner (MFA) tightens word boundaries.

The released corpus is split into short, role-consistent segments. Each segment carries sociodemographic metadata and geocoded school coordinates. On manually annotated data, the LLM step raised turn-level role attribution accuracy from 62.70% to 95.68%.

The resource opens historical Australian English to phonetic, lexical, grammatical, discourse and spatial analysis. The author also presents the pipeline as reusable for other interview-style legacy dialect archives.

Key points

  • The source is the MD collection: 7,735 pupils from 327 schools across all Australian states and territories, whose recordings were made in 1959/60 and later converted to digital form by the University of Sydney in 1998.
  • The tools are WhisperX with the Whisper large-v3 model, Pyannote's speaker-diarization-community-1 model (set to 1–2 speakers), Gemini for discourse-role correction, and MFA.
  • Speaker-role attribution accuracy rose from 62.70% with acoustic diarization alone to 95.68% after adding LLM reasoning.
  • MFA runs on windows of at most 12 tokens. The word forms always stay the WhisperX tokens, and MFA only replaces timestamps when its alignment succeeds.
  • The metadata covers sex, birthplace, parental background, school location and geocoordinates, which supports spatially explicit dialectology.

Interesting ideas

  • I1. A text-level reasoning step uses conversational structure, such as question–answer pairing, to fix role labels that a purely acoustic system got wrong.
  • I2. One tool is the fixed authority for token identity, while a second tool may only adjust one attribute (timing), and only where it succeeds.
  • I3. The speaker-count setting allows anywhere from one to two speakers, because some recordings have the interviewer faint, absent or erased.
  • I4. Text is normalized to match the aligner's dictionary before alignment, for example expanding "50%" into "50 percent".
  • I5. The LLM step is evaluated against the acoustic-only baseline on hand-annotated data, which isolates the value the extra stage adds.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (frames): question/answer speech-act detection in prag could help assign or verify who is speaking in dialogues from fables and tales, beyond surface cues.
  • I2 (learn): when building silver data, freeze the en tokenization and take only arcs or labels from a witness parser that aligns cleanly, falling back to the original otherwise.
  • I3 (world): entity tracking in a text could allow for an implied but unvoiced participant (an unseen questioner) instead of forcing a fixed cast.
  • I4 (tokens): a normalization layer that expands symbols and numbers before lemma or dictionary lookup, while keeping a mapping to the original tokens, would reduce lookup failures.
  • I5 (learn): run each LLM teacher tier as an ablation against the plain pipeline on a small gold set, so its contribution is measured rather than assumed.

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

🎓University of OuluOulu · Finland2 works · 38 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 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksEnglish · 2,263 works

Cite

Cite as Steven Coats. 2026. MD_NLP: Reconstructing an Australian English Heritage Dialect Corpus from the Mitchell-Delbridge Recordings through LLM-Assisted Speaker Attribution. Proceedings of the First Workshop on Dialects in NLP –- A Resource Perspective.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3iga3zzsh92p

@inproceedings{coats2026md,
  title = {{MD\_NLP: Reconstructing an Australian English Heritage Dialect Corpus from the Mitchell-Delbridge Recordings through LLM-Assisted Speaker Attribution}},
  author = {Steven Coats},
  booktitle = {Proceedings of the First Workshop on Dialects in NLP –- A Resource Perspective},
  year = {2026},
  doi = {10.63317/3iga3zzsh92p},
  url = {https://doi.org/10.63317/3iga3zzsh92p}
}

References

21 references, 3 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. In Interspeech 2023, pages 2573– 2577. find
  2. Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations.
  3. Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. 2023. WhisperX: Time-accurate speech transcription of long-form audio. In IN-TERSPEECH 2023, pages 4489–4493. find
  4. Hervé Bredin. 2023. pyannote.audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe. In INTERSPEECH 2023, pages 1983– 1987. find
  5. Luyao Cheng, Hui Wang, Chong Deng, Siqi Zheng, Yafeng Chen, Rongjie Huang, Qinglin Zhang, Qian Chen, Xihao Li, and Wen Wang. 2025. Integrating audio, visual, and semantic information for enhanced multimodal speaker diarization on multi-party conversation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 19914–19928, Vienna, Austria. Association for Computational Linguistics. ↗
  6. Steven Coats, Carmelo Alessandro Basile, Cameron Morin, and Robert Fuchs. 2025a. The YouTube corpus of Singapore English podcasts. English World-Wide, 46(3):274–298. ↗
  7. Steven Coats, Chloé Diskin-Holdaway, and Debbie Loakes. 2025b. Regional Distribution of the /el/-/æl/ Merger in Australian English. In Proceedings of the 12th Workshop on NLP for Similar Languages, Varieties and Dialects, pages 147–156, Abu Dhabi, UAE. Association for Computational Linguistics.
  8. Rolando Coto-Solano, James N. Stanford, and Sravana K. Reddy. 2021. Advances in completely automated vowel analysis for sociophonetics: Using end-to-end speech recognition systems With DARLA. Frontiers in Artificial Intelligence, 4. ↗
  9. Felicity Cox and Sallyanne Palethorpe. 2019. Vowel variation in a standard context across four major Australian cities. In Proceedings of the 19th International Congress of Phonetic Sciences, Melbourne, Australia 2019, pages 577–581. Australasian Speech Science and Technology Association. find
  10. Felicity Cox, Sallyanne Palethorpe, and Samantha Bentink. 2014. Phonetic archaeology and 50 years of change to Australian English /i:/. Australian Journal of Linguistics, 34(1):50–75. find
  11. Felicity Cox, Joshua Penney, and Sallyanne Palethorpe. 2024. Australian English monophthong change across 50 Years: Static versus dynamic measures. Languages, 9(3). ↗
  12. Eric Harper, Somshubra Majumdar, Oleksii Kuchaiev, Jason Li, Yang Zhang, Evelina Bakhturina, Vahid Noroozi, Sandeep Subramanian, Nithin Koluguri, Jocelyn Huang, Jia Fei, Jagadeesh Balam, Xuesong Yang, Micha Livne, Yi Dong, Naren Sean, and Boris Ginsburg. 2023. NeMo: a toolkit for Conversational AI and Large Language Models. find
  13. Guillaume Klein and Mahmoud Ashraf. 2023. faster-whisper: Faster whisper transcription with ctranslate2. find
  14. Nikola Ljubešić, Peter Rupnik, and Tea Perinčić. 2024. Mići Princ – A Little Boy Teaching Speech Technologies the Chakavian Dialect. In Proceedings of the Conference on Language Technologies and Digital Humanities, pages 232–250. Institute of Contemporary History. ↗
  15. Deborah Loakes, John Hajek, and Janet Fletcher. 2017. Can you t[æ]ll I’m from M[æ]lbourne?: An overview of the dress and trap vowels before /l/ as a regional accent marker in Australian English. English World-Wide. A Journal of Varieties of English, 38(1):29–49. find
  16. Michael McAuliffe, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger. 2017. Montreal Forced Aligner: Trainable text-speech alignment using Kaldi. In Interspeech 2017, pages 498–502. find
  17. Alexander George Mitchell and Arthur Delbridge. 1998. The speech of Australian adolescents: Research data and recordings collected by AG Mitchell and Arthur Delbridge in 1959 and 1960. The University of Sydney. find
  18. Alexis Plaquet and Hervé Bredin. 2023. Powerset multi-class cross entropy loss for neural speaker diarization. In INTERSPEECH 2023, pages 3222–3226. find
  19. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 28448–28481. PMLR. ↗
  20. Penelope Schmidt, Chloé Diskin-Holdaway, and Debbie Loakes. 2021. New insights into /el/-/æl/ merging in Australian English. Australian Journal of Linguistics, 41(1):66–95. _eprint: https://doi.org/10.1080/07268602.2021.1905607. ↗
  21. Quan Wang, Yiling Huang, Guanlong Zhao, Evan Clark, Wei Xia, and Hank Liao. 2024. DiarizationLM: Speaker Diarization Post-Processing with Large Language Models. In Interspeech 2024, pages 3754–3758. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/coats-2026-md-nlp-reconstructing-australian-english/