Perceptual compensation for tonal context in self-supervised speech models

In three linesStudy of wav2vec2.0 examining perceptual compensation for tonal context in Mandarin Chinese. Purely self-supervised pre-trained models show no evidence of compensation in embedding similarities. Probing classifiers reveal partial compensation but fail to replicate human performance. Supervised objectives appear necessary to abstract certain phonological regularities.

Summary generated by Claude — human-verified