Reproduce “Easper: An Accessible ASR Pipeline for Language Documentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aso Mahmudi, Ting Dang, Ekaterina Vylomova, Nick Thieberger
The University of Melbourne developed Easper, an accessible, cloud-based, no-code ASR pipeline for language documentation that integrates with ELAN and provides empirical guidelines for data selection. Their work demonstrated that prioritizing training data with high Normalized Token-to-Type Ratio (ToTy) leads to faster improvements in ASR transcription quality for low-resource languages, outperforming strategies based on acoustic cleanliness.
Key insights
- Linguistic characteristics of training data, particularly lexical repetition, are more critical than acoustic cleanliness for rapidly bootstrapping ASR models in low-resource settings.
- Large pre-trained foundation models like Whisper demonstrate inherent robustness to common field recording issues such as background noise and speaker overlap.
- Neural networks require sufficient acoustic-phonetic repetitions to reliably learn new grapheme-to-phoneme mappings, making sessions rich in word repetition (high ToTy) more effective for early fine-tuning than those with broad but sparse vocabulary (high TyTo).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.