A Preliminary Study on Automatic Pronunciation Error Detection for Hearing-impaired Children
Yingming Gao, Hai Shuang, Xiaoli Feng, Jingwen Cheng, Linkai Peng, Ya Li, Jinsong Zhang, Min Liu · 2024
Due to limited economic resources and a lack of specialized educational support, most hearing-impaired children are unable to access basic education or specialized rehabilitation services.Therefore, it is important to study their language acquisition patterns and to develop computer-assisted language learning (CALL) systems that can help them return to normal life and basic education as much as possible.However, previous work has mainly focused on the descriptive analysis of speech production, and current CALL systems are also mainly aimed at language learning for people with normal language abilities.Therefore, in this paper, we first constructed a Chinese speech dataset (60 hearing-impaired and 107 normal-hearing children) containing 294 monosyllabic or disyllabic words.Additionally, we recruited 10 trained annotators to manually label phones and their boundaries.Furthermore, we developed pronunciation error detection systems based on an end-to-end speech recognition architecture and conducted a preliminary exploration of the performance of different modeling units.The experimental results show that hearing-impaired children have difficulties in speech production, and the correct rates of consonants, vowels, and tones are 63.8%, 76.9%, and 60.59%, respectively.The pronunciation error detection systems achieved recognition error rates of 69.21%, 49.75%, and 61.77% for characters, phones, and tones, respectively.