NLPP: A Natural Language Prosodic Prominence Dataset Assisted by ChatGPT

Recent advancements in speech synthesis, leveraging deep learning and vast datasets, have largely addressed accuracy and fluency concerns. However, expressive speech synthesis, aiming for nuanced styles like emotion and pace, remains a challenge. Prior methods relying on categorical labels or short prompts limit authenticity. Addressing this, recent approaches incorporate natural language prompts for flexibility and ease of use. In this paper, we present NLPP, a novel dataset describing prosodic prominence through natural language with the assisting of ChatGPT. The experimental results demonstrate the rich-ness and accuracy of NLPP in description. NLPP pioneers natural language descriptions of prosodic prominence, facilitating more authentic expressive speech synthesis. Meanwhile, NLPP also serve as a processing paradigm for related tasks.