MD_NLP

Dataset Description

MD_NLP is a discourse-annotated, word-aligned, and georeferenced corpus derived from the narrative portion of the Mitchell–Delbridge recordings, a large mid-20th-century archive of Australian English. The corpus was constructed from archival WAV recordings using an automated pipeline combining WhisperX-based ASR, neural speaker diarization, LLM-assisted discourse-role correction, and Montreal Forced Aligner boundary refinement.

The released dataset consists of short, role-consistent narrative segments with transcripts, word-level timestamps, linked audio, and selected sociodemographic metadata.

• Curated by: Steven Coats
• Institution: University of Oulu
• Language(s): English (Australian English)
• License: cc-by-nc-4.0
• Related paper: MD_NLP: Reconstructing an Australian English Heritage Dialect Corpus from the Mitchell–Delbridge Recordings through LLM-Assisted Speaker Attribution

Dataset Summary

The source archive comprises recordings of 7,735 Australian secondary school pupils from 327 locations across Australia, recorded in 1959–1960. MD_NLP includes the spontaneous narrative component of these recordings rather than the read word-list and sentence materials more commonly used in previous research.

The dataset is intended for research on:

• Australian English variation
• dialectology and sociolinguistics
• discourse structure and turn-taking
• corpus phonetics
• ASR, diarization, and alignment on legacy speech recordings

Dataset Structure

Each row corresponds to a short, role-consistent segment.

Fields

• interview_id: numeric interview identifier
• segment_id: unique segment identifier
• role: discourse role label (Student or Teacher)
• start: segment start time in seconds
• end: segment end time in seconds
• transcript: transcript text for the segment
• word_tokens: list of word-level tokens with start and end times
• audio: path/reference to the corresponding audio segment
• student_sex: recorded sex metadata for the student
• state: Australian state or territory
• town_city: locality
• recording_year: recording year
• institution: school/institution name

Split

The current release contains one split:

• train: 257,357 segments

Corpus Size

• Recording length: 214.14 hours
• Speech duration: 137.95 hours
• Turns: 71,929
• Word count: 1,791,856

Role-based totals:

Metric	Student	Teacher	Total
Speech duration (h)	92.71	45.24	137.95
Turns	46,026	25,903	71,929
Word count	1,155,994	635,862	1,791,856

Source Data

Mitchell, Alexander George 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. https://doi.org/10.25910/jkwy-wk76

The dataset is derived from the Mitchell–Delbridge recordings, made by schoolteacher volunteers in 1959 and 1960 in 327 locations across all Australian states and territories. The original archive contains read materials and a short narrative component. MD_NLP includes only the narrative recordings.

The narratives typically involve brief teacher–student interaction, though some recordings are more monologic. Recording conditions vary substantially across sites.

Processing

The corpus was created using the following pipeline:

1. WhisperX for automatic speech recognition and initial word alignment
2. Pyannote for speaker diarization
3. LLM-assisted discourse-role correction (Gemini 2.5-flash) to distinguish Teacher and Student
4. Montreal Forced Aligner (MFA) for boundary refinement
5. Reconstruction into short, role-consistent segments with word-level timing

The released transcripts preserve the original WhisperX tokenization while using refined timestamps where alignment succeeded.

Evaluation

Speaker-role attribution was evaluated on 10 manually checked narratives (approximately 30 minutes of speech; 185 turns).

System	Accuracy
Baseline (WhisperX + Pyannote)	62.70%
Full pipeline (LLM-assisted)	95.68%

These results indicate that the LLM-assisted role-correction step substantially improves turn-level speaker attribution in interview-style archival recordings.

Intended Use

MD_NLP is intended for research use, especially for:

• regional and social variation in Australian English
• discourse and interactional structure
• corpus phonetics and time-aligned speech analysis
• geographically explicit dialect research
• evaluation of ASR, diarization, and alignment methods on legacy speech

Limitations

• The corpus is derived from archival recordings with variable audio quality.
• Some interviewer speech is faint, partially absent, or missing.
• Transcripts are automatically generated and corrected, not manually transcribed throughout.
• Some alignment boundaries may remain imperfect despite MFA refinement.
• Metadata reflect archival source records and may contain inconsistencies or omissions.

Sensitive Information

The dataset contains speech-derived transcripts and linked metadata fields such as sex, institution, state, town/city, and recording year. These are historical archival data. Users should handle the dataset in accordance with the license and any archive-specific restrictions.

Citation

If you use this dataset, please cite the associated paper.

BibTeX

@inproceedings{coats2026mdnlp,
  title={MD\_NLP: Reconstructing an Australian English Heritage Dialect Corpus from the Mitchell--Delbridge Recordings through LLM-Assisted Speaker Attribution},
  author={Coats, Steven},
  booktitle={Proceedings of LREC 2026},
  year={2026}
}