A Wav2Vec2 and Multi-Head Attention Based Framework for Pronunciation Error Detection in Tibetan Mandarin Learners

Abstract

Aiming at the systematic pronunciation bias problem of Tibetan native speakers when learning Mandarin Chinese, this paper proposes an end-to-end detection method based on the self-supervised learning Wav2Vec 2.0 model fusing the multi-head self-attention mechanism (MHA) with a CTC decoder. The model is fine-tuned to adapt to the pronunciation characteristics of Tibetan, and the MHA is utilized to enhance the ability of capturing long-distance dependent features. Experiments on a self-constructed Mandarin pronunciation bias dataset of Tibetan students show that the proposed model significantly outperforms the traditional ASR model and the baseline Wav2Vec2-CTC system in terms of detection accuracy (DAR) and F1 scores, which validates its effectiveness in low-resource speech learning scenarios.