← Research & recognition

A Novel Data Augmentation Approach for Automatic Speaking Assessment on Opinion Expressions

Authors: Chung-Chun Wang, Jhen-Ke Lin, Hao-Chien Lu, Hong-Yun Lin, Berlin Chen

Two people beside the poster titled “A Novel Data Augmentation Approach for Automatic Speaking Assessment on Opinion Expressions,” from the SLaTE 2025 presentation post.

Two people beside the poster titled “A Novel Data Augmentation Approach for Automatic Speaking Assessment on Opinion Expressions,” from the SLaTE 2025 presentation post. · Photo source

Publication: 10th Workshop on Speech and Language Technology in Education (SLaTE 2025), pp. 199–203

My role: First author.

Problem: Speaking assessment on opinion expressions lacks labeled recordings, which limits prompt diversity and undermines scoring reliability.

Method: An LLM generates diverse responses at a given proficiency level, speaker-aware TTS synthesizes them into speech, and a dynamic importance loss reweights training instances by the feature-distribution gap between synthesized and real speech; a multimodal LLM then combines text and speech features to predict proficiency scores directly.

Result: On the LTTC dataset, the approach outperforms methods relying on real data or conventional augmentation, easing the low-resource constraint.