Tone-Aware Transformer Model for Chinese Dialect Recognition and Preservation

Most AI language tools struggle with regional Chinese dialects, understanding only textbook Mandarin. DialectBERT solves this by preserving linguistic diversity, ensuring every speaker is heard—not just those using Standard Mandarin. Unlike traditional systems, it learns nuanced tonal variations ("mā" vs. "mà") like a skilled linguist, mastering both common and low-resource dialects while handling everyday code-switching. The results are groundbreaking: DialectBERT achieves 93.5% accuracy in dialect recognition, surpassing earlier systems stuck at 85%, with precision (92.8%) and recall (94.0%) equally strong. This isn’t just incremental improvement it’s a leap toward inclusivity for Shanghainese elders, Hakka merchants, and others often excluded by tech. Beyond metrics, the project embodies responsible AI, prioritizing linguistic diversity over assimilation. Efforts are underway to deploy DialectBERT on affordable smartphones for rural communities and collaborate with anthropologists to document endangered dialects. The goal? Technology that adapts to people’s native languages—not the other way around.