Contrastive Pretraining Hits 91.6% Syllable-Level Tone Classification at 10 Shots, and Transfers From Vietnamese to Mandarin

arXiv·low signalXBlueskyLinkedInCopy link

ToneCL (arXiv 2609.24903, 21 Sep 2026) targets a gap field linguists actually hit: tone languages are 50-70% of the world's languages and mostly low-resource, and existing datasets are sentence-level where the annotation needed is syllable-level. The framework pretrains on unlabeled speech with augmentations that preserve tonal identity, then fine-tunes on tens of examples per tone class, reaching 91.6% on six-speaker Mandarin at 10 shots. Cross-lingual transfer works too, with Vietnamese pretraining then Mandarin fine-tuning reaching 91.0% at 10 shots, and the ablation names frequency band rejection as the critical augmentation.