Pretrained self-supervised speech models can recognize unseen consonants

In three linesPretrained self-supervised speech models (Wav2Vec2, HuBERT) fine-tuned on Khoisan languages (G|ui, West !Xoon) recognize click consonants more accurately than non-clicks, showing self-supervision generalizes to rare phonemes despite training data skewed toward high-resource languages.

Summary generated by Claude — human-verified