MD_NLP: Reconstructing an Australian English Heritage Dialect Corpus from the Mitchell-Delbridge Recordings through LLM-Assisted Speaker Attribution
Year: 2026
Type: conference-paper
Abstract: We present MD_NLP, a discourse-annotated and georeferenced corpus derived from the Mitchell-Delbridge (MD) recordings, a foundational archive of mid-20th-century Australian English.The corpus comprises word-aligned narrative recordings from 7,735 secondary school pupils across 327 locations, enriched with structured sociodemographic metadata (e.g., sex, birthplace, parental background) and geocoded institutional coordinates.The narratives were reconstructed from archival audio using an integrated pipeline combining WhisperX-based automatic speech recognition, neural speaker diarization, LLM-assisted discourse-role correction, and Montreal Forced Aligner (MFA) boundary refinement.Evaluation on manually annotated data shows that incorporating an LLM-based reasoning step improves turn-level speaker-role attribution from 62.70% (acoustic diarization alone) to 95.68%.Unlike prior uses of the MD archive, which focused on controlled sentence materials, MD_NLP makes the spontaneous narrative component accessible for large-scale analysis.The resulting resource supports research on regional and socially conditioned variation, discourse structure, and corpus phonetics in Australian English.The proposed architecture is directly transferable to other legacy dialect archives, providing a practical pathway for transforming interview-based recordings into temporally aligned, speaker-consistent corpora.
Author: Steven Coats
Institution: University of Oulu
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166589580