Elamur Research⌕
☰
home / papers / 2022 / The Corpus of Australian and New Zealand Spoken English: A new resource of naturalistic speech transcripts

The Corpus of Australian and New Zealand Spoken English: A new resource of naturalistic speech transcripts

Paper · 2022 · Steven Coats

Summary

Studies of regional grammar in contemporary spoken English have had little suitable data. Existing Australian and New Zealand speech corpora tend to be small, concentrated in a few cities, or made up largely of older recordings, so rare syntactic features and regional patterns are hard to study. Coats addresses this with CoANZSE. He compiled official lists of local government bodies, scraped their websites for YouTube links and manually confirmed each channel. Each channel was geocoded from its street address. He then downloaded the automatic captions with yt-dlp, added word timings and POS-tagged them with spaCy. The result is about 190 million words of naturalistic, geolocated, time-aligned speech. Two pilot studies show what the corpus can do. A search for double modals finds genuine Australian uses of a construction usually described as absent there. A simple classifier also separates Australian from New Zealand transcripts reasonably well using only common words. Because every token links back to its video, the resource supports error checking and phonetic or multimodal follow-up work.

Key points

  • CoANZSE's 56,815 ASR transcripts amount to about 195.5M tokens and approximately 24,030 video hours, drawn from 482 YouTube channels operated by councils and similar bodies in 472 locations. These channels come from 696 listed entities, 578 of which had websites, yielding 515 channels found.
  • The pipeline uses yt-dlp to retrieve captions, a custom parser that keeps per-word timing, spaCy en_core_web_sm for POS tags, and geocoding of council addresses.
  • The double-modal query returned 3,119 candidates. Around 400 of these were checked against the videos, and many proved to be authentic Australian uses.
  • The country classifier used tf-idf over the 500 most frequent words and a linear SVM tuned with GridSearchCV and balanced class weights, trained on 80% of a 10,000-transcript sample. Accuracy was 0.80, with F1 of 0.86 for AU and 0.66 for NZ (NZ recall 0.59).
  • Where YouTube also offers human-made captions, the measured ASR word error rate is about 0.14. Transcripts carry no speaker diarization.

Interesting ideas

  • I1. Wherever the same videos have both machine and human captions, the paper compares the two to get an error estimate cheaply, without commissioning new annotation.
  • I2. Rare phenomena are handled in two steps: a broad, high-recall search over the noisy data, followed by fast human checking. The checking is quick because every hit links straight to its source moment.
  • I3. The author advises relying on frequent features and coarse aggregation when data is noisy, since correct cases then outweigh transcription errors, and reserving manual checks for rare features.
  • I4. The paper treats a non-standard construction like "might could" as a question of frequency, not as something categorically absent from a variety.
  • I5. Because the transcripts lack speaker turns, analyses that depend on who is speaking cannot use them directly without further annotation.

For Mova

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

  • I1 (learn): Find the overlap between silver parses and existing gold UD annotation, and use that overlap to measure the error rate of each witness parser before trusting it.
  • I2 (learn): Silver-data mining for rare patterns should keep a pointer from each hit back to its sentence and parse, so that an LLM teacher or a human can confirm or reject a sample quickly.
  • I3 (learn): Automatically induced repair rules could be restricted to high-frequency patterns, with low-count patterns sent to verification instead of being learned directly.
  • I4 (tree): The parser and tagger should be able to attach stacked modals as consecutive aux dependents instead of forcing an error, and a few verified double-modal sentences could serve as a regression probe.
  • I5 (frames): When Mova reads undiarized dialogue-like text, first-person pronouns should be marked as having an unresolved speaker instead of being merged into a single referent during coreference.

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, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

🧭 Topics 1Research subjects and languages the work is about.English · 2,263 works

Cite

Cite as Steven Coats. 2022. The Corpus of Australian and New Zealand Spoken English: A new resource of naturalistic speech transcripts. Proceedings of the 20th Annual Workshop of the Australasian Language Technology Association.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{coats2022corpus,
  title = {{The Corpus of Australian and New Zealand Spoken English: A new resource of naturalistic speech transcripts}},
  author = {Steven Coats},
  booktitle = {Proceedings of the 20th Annual Workshop of the Australasian Language Technology Association},
  year = {2022},
  url = {https://aclanthology.org/2022.alta-1.1/}
}

References

25 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sumeet Agarwal, Shantanu Godbole, Diwakar Punjani, and Shourya Roy. 2007. How much noise is too much: A study in automatic text classification. In Seventh IEEE International Conference on Data Mining (ICDM 2007), pages 3–12. find
  2. Alëna Aksënova, Daan van Esch, James Flynn, and Pavel Golik. 2021. How might we create better benchmarks for speech recognition? In Proceedings of the 1st Workshop on Benchmarking: Past, Present and Future, pages 22–34, Online. Association for Computational Linguistics. find
  3. Liselotte Anderwald and Suzanne Wagner. 2007. The Freiburg English Dialect Corpus: Applying corpus-linguistic research tools to the analysis of dialect data. In Joan C. Beal, Karen P. Corrigan, and Hermann Moisl, editors, Creating and digitizing language corpora volume 1: Synchronic databases, pages 35–53. Palgrave Macmillan, Houndmills, Basingstoke. find
  4. Julie Bradshaw, Kate Burridge, and Michael Clyne. 2010. The Monash Corpus of Spoken Australian English. In Proceedings of the 2008 Conference of the Australian Linguistics Society. Australian Linguistic Society. find
  5. Kate Burridge. 2020. History of Australian English. In Louisa Willoughby and Howard Manns, editors, Australian English reimagined: Structure, features and developments, pages 175–192. Routledge. ↗
  6. Steve Cassidy, Michael Haugh, Pam Peters, and Mark Fallu. 2012. The Australian national corpus: National infrastructure for language resources. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), pages 3295–3299, Istanbul, Turkey. European Language Resources Association (ELRA). find
  7. Steven Coats. 2022. Naturalistic double modals in North America. American Speech. Steven Coats. In review. Double Modals in contemporary British and Irish Speech. ↗
  8. Karen P. Corrigan, Isabelle Buchstaller, Adam Mearns, and Hermann Moisl. 2012. The Diachronic Electronic Corpus of Tyneside English. ↗
  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. John W. Du Bois, Wallace L. Chafe, Charles Meyer, Sandra A. Thompson, Robert Englebretson, and Nii Martey. 2000-2005. Santa Barbara Corpus of Spoken American English. find
  11. Jonathan Dunn. 2019. Global syntactic variation in seven languages: Toward a computational dialectology. Frontiers in Artificial Intelligence, Section Language and Computation. ↗
  12. Elizabeth Gordon, Margaret Maclagan, and Jennifer Hay. 2007. The ONZE corpus. In Joan C. Beal, Karen P. Corrigan, and Hermann Moisl, editors, Creating and digitizing language corpora volume 2: Diachronic databases, pages 82–104. Palgrave Macmillan, Houndmills, Basingstoke. ↗
  13. Sidney Greenbaum, editor. 1996. Comparing English worldwide: The International Corpus of English. Clarendon Press. find
  14. Sidney Greenbaum. 1998. A proposal for an international computerized corpus of english. World Englishes, 7(3):315. find
  15. Jack Grieve, Chris Montgomery, Andrea Nini, Akira Murakami, and Diansheng Guo. 2019. Mapping lexical dialect variation in British English using Twitter. Frontiers in Artificial Intelligence, 2. ↗
  16. Michael Haugh and Wei-Lin Melody Chang. 2013. Collaborative creation of spoken language corpora. In Tim Greer, Yuriko Kite, and Donna Tatsuki, editors, Pragmatics and Language Learning, Volume 13, pages 133–159. National Foreign Language Resource Center, University of Hawaii. find
  17. Janet Holmes, Bernadette Vine, and Gary Johnson. 1998. Guide to the Wellington Corpus of Spoken New Zealand English. find
  18. Dirk Hovy and Christoph Purschke. 2018. Capturing regional variation with distributed place representations and geographic retrofitting. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4383–4394. find
  19. Thorsten Joachims. 1998. Text categorization with support vector machines: Learning with many relevant features. Springer. ↗
  20. Bernd Kortmann. 2010. Areal variation in syntax. In Peter Auer and Jürgen E. Schmidt, editors, Language and space: An international handbook of linguistic variation, volume 1, theories and methods, pages 837–64. de Gruyter Mouton. find
  21. Bernd Kortmann and Kerstin Lunkenheimer, editors. 2013. The Electronic World Atlas of Varieties of English (eWAVE). Max Planck Institute for Evolutionary Anthropology, Leipzig. find
  22. Lee Murray and Howard Manns. 2020. Lexical and morphosyntactic variation in australian english. In Louisa Willoughby and Howard Manns, editors, Australian English reimagined: Structure, features and developments, pages 120–133. Routledge. ↗
  23. John Nerbonne. 2009. Data-driven dialectology. Language and Linguistics Compass, 3:175–198. ↗
  24. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830. ↗
  25. Benedikt Szmrecsanyi. 2011. Corpus-based dialectometry: A methodological sketch. Corpora, 6(1):45–76. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/coats-2022-corpus-australian-new-zealand-spoken/