A new contrastive learning framework adapts a speech-trained HuBERT model to unlabeled violin recordings, achieving the best ...