LLM-Based Multi-Reference Evaluation for Efficient and Robust Assessment of Phrase Break Annotations

The article presents the LLM-based Multi-Reference Evaluation (LMRE) framework designed for assessing phrase break annotations in speech, addressing the limitations of single-reference evaluation methods. LMRE utilizes large language models to generate multiple valid phrasing options from minimal demonstrations, demonstrating improved alignment with human judgment on a Korean testbed of 1,356 annotations. This approach offers a scalable and robust evaluation method that leverages the one-to-many nature of prosodic phrasing, which is significant for practitioners focusing on speech clarity and naturalness in AI applications.