Perceptual compensation for tonal context in self-supervised speech models

A new study examining the wav2vec2.0 self-supervised speech model found no evidence that it spontaneously learns to compensate for phonological context in Mandarin Chinese tones, challenging assumptions about what linguistic structure emerges from pre-training alone [1]. The research, submitted to the arXiv preprint repository on June 16, 2026, conducted a pseudo-replication of a human perceptual compensation experiment [1]. The authors compared embedding similarities and probing classifier outputs between a purely self-supervised pre-trained model and one fine-tuned for Mandarin automatic speech recognition [1]. No evidence of compensation was found in the embedding similarities of the purely pre-trained model [1]. Probing classifiers showed some evidence of compensation alongside expected layer-wise improvements in categorization, but the classifiers failed to replicate human performance on isolated test syllables [1]. The findings contrast with earlier reports suggesting that sensitivity to phonological structure could emerge through pre-training alone [1]. The authors suggest that supervised objectives may be necessary to encourage the abstraction of at least some types of phonological regularities [1]. arXiv, where the paper appeared, is an open-access repository of electronic preprints that are moderated but not peer-reviewed [6]. Founded in 1991, the repository passed the two-million-article milestone by the end of 2021 and now receives about 24,000 submissions per month as of November 2024 [6]. The platform hosts papers across mathematics, physics, computer science, and related fields [6]. The study falls within the Computation and Language category of computer science [1]. The wav2vec2.0 architecture examined in the study is a type of machine learning model trained with self-supervised learning on large amounts of speech data [8]. Self-supervised models learn representations without explicit labels, and researchers have been investigating whether such models internalize abstract linguistic structures [1]. The new results indicate that at least for tonal context compensation — a perceptual phenomenon where listeners adjust their interpretation of tones based on surrounding speech sounds — the purely self-supervised model did not develop human-like behavior [1]. The fine-tuned model showed some improvement, but still fell short of human performance on isolated syllables [1]. The paper is available through arXiv's standard interface, which includes experimental community tools developed under the arXivLabs framework [4]. arXivLabs, launched in 2020, allows collaborators to build features such as bibliographic explorers and recommender systems that integrate directly with the article page [4]. These tools operate under guidelines that require partners to share arXiv's values of openness, community, excellence, and user data privacy [4].