Low-Resource NLP for Indigenous Language Preservation using Speech Transcription and Lexical Resource Creation

Abstract

Indigenous languages maintain ecological knowledge, oral literature, identity and memory of the community, but many do not have transcribed speech, normalized orthographies, and computer-readable dictionaries for contemporary language technologies. This paper presents a low-resource natural language processing framework, which combines speech transcription and lexical resource building. The framework considers automatic speech recognition as a redrafting machine that can be corrected by a human, matches corrected transcripts to dictionary entries within the candidate's dictionary, accommodates orthographic variations and stores dictionary metadata that has been reviewed by the culture. The method consists of community-based data governance, multilingual speech representation transfer, confidence-based human review, OCR-assisted archival extraction and lexicon-aware normalization. Results obtained from a simulation-oriented evaluation show reduced WERs, higher dictionary coverage and lower manual transcription time than those obtained from a manual-only and ASR-only baselines. The proposed design focuses on the transparency of the review process, dialect preservation and the reuse of digital outputs for education, documentation and revitalization.