Tag
A ten-seed study of self-supervised speech encoders shows that the learning objective (reconstruction vs. prediction) governs cross-lingual perceptual narrowing, with reconstruction degrading non-native phoneme discrimination and prediction improving it.