Acoustic Unit-Based System Empowering Accent Correction Across Low-Resource Languages

Mainstream accent correction systems still require labeled data during training, which limits the technology to a small fraction of global languages. Particularly challenging is the fact that some languages lack a written form, preventing the generation of necessary textual annotations and thereby blocking the feasibility of traditional supervised learning approaches. Moreover, even existing L2 speech features often do not perfectly match their corresponding textual annotations. These issues highlight the urgency of exploring unsupervised learning methods in the field of accent correction, especially when dealing with low-resource languages and challenging L2 speech feature modeling. Therefore, we propose the Unsupervised Acoustic Unit-based Accent Correction Model (UAU-ACM), which detects and modifies accentuated speech using an unsupervised acoustic unit approach. Experiments conducted on the public KeSpeech dataset comprehensively validate the superiority of our model.