Reproduce “Evaluating Self-supervised Speech Models on a Taiwanese Hokkien Corpus” 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
Yi-Hui Chou, Kalvin Chang, Meng-Ju Wu, Winston Ou, Alice Wen-Hsin Bi, Carol Yang, Bryan Y. Chen, Rong-Wei Pai and 7 more
Researchers from Carnegie Mellon University and several Taiwanese institutions developed and contributed a 1.5-hour Taiwanese Hokkien speech corpus to ML-SUPERB for Automatic Speech Recognition (ASR) evaluation. Their study revealed that self-supervised models pretrained on linguistically similar Mandarin data achieved superior performance over larger multilingual models, identifying tone prediction as a primary source of ASR error, responsible for a 16.6% relative increase in Syllable Error Rate.
Key insights
- Linguistic alignment of pretraining data is more critical for robust ASR in low-resource settings than the sheer quantity or diversity of the pretraining corpus, especially when a related language is available.
- Tone prediction presents a substantial challenge for Taiwanese Hokkien ASR, significantly hindering model performance and warranting specialized research focus.
- Larger model sizes do not guarantee improved performance for low-resource languages; the efficacy of scaling depends on the specific pretraining strategy and data characteristics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.