Personalized Learning of Tibetan Academic Mandarin Pronunciation Integrating Mask-GCT Voice CloningTechnology

Abstract

This study proposes a personalized Mandarin pronunciation learning framework for Tibetan learners that integrates Mask‑GCT voice‑cloning technology as a back‑end data‑augmentation module. By leveraging deep neural networks, the voice‑cloning component reconstructs key speaker characteristics—timbre, intonation, and prosody—from limited samples, generating high‑fidelity, individualized speech data. These synthetic samples not only alleviate labeled‑data scarcity but also introduce diverse pronunciation scenarios, particularly modeling the tonal, vowel, and consonant errors characteristic of Tibetan students’ Mandarin production. Adjustable cloning parameters enable simulation of multiple error patterns and subtle phonetic variations, thereby enriching training data and enhancing the model’s capacity to detect, adapt to, and correct a wide range of pronunciation deviations. Experimental results demonstrate that our approach significantly improves error recognition accuracy and model generalization compared to baseline systems lacking voice‑cloning augmentation. The flexible, controllable synthesis process provides empirical support for targeted pronunciation remediation, offering a scalable methodology for assisting Tibetan learners in mastering academic Mandarin pronunciation.