Pretrained self-supervised speech models can recognize unseen consonants

The study fine-tunes pretrained self-supervised speech models, specifically Wav2Vec2 and HuBERT, to evaluate their ability to recognize click consonants from Khoisan languages. Results indicate that these models achieve higher accuracy in recognizing click consonants compared to non-click sounds, demonstrating their potential for generalizing across diverse phonetic inventories. This advancement is significant for practitioners as it highlights the effectiveness of self-supervised learning in addressing underrepresentation of low-resource languages and rare phonemes in speech recognition tasks.