Combining LLM-generated text, speaker-cloned TTS speech, and a dynamic importance loss improves automated speaking assessment accuracy by 3.3% on seen prompts and under 1% on unseen prompts versus real-data-only training on the LTTC dataset.
Towards automatic scoring of a test of spoken language with heterogeneous task types,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
A Novel Data Augmentation Approach for Automatic Speaking Assessment on Opinion Expressions
Combining LLM-generated text, speaker-cloned TTS speech, and a dynamic importance loss improves automated speaking assessment accuracy by 3.3% on seen prompts and under 1% on unseen prompts versus real-data-only training on the LTTC dataset.