MELT
收藏资源简介:
MELT数据集是基于美剧《老友记》的多模态情绪数据集,由GPT-4o模型自动标注而成。该数据集包含来自《老友记》的对话片段,共计8821条语句,涵盖了7种情绪类别。MELT数据集的创建过程主要涉及了对话片段的筛选、GPT-4o模型的选取和提示工程设计等方面。该数据集旨在解决语音情绪识别(SER)中标注成本高、一致性差的问题,并通过主观和客观实验验证了其标注质量和模型性能的提升。
The MELT dataset is a multimodal emotion dataset based on the American TV series *Friends*, automatically annotated by the GPT-4o model. It contains 8,821 utterances from the dialogue segments of *Friends*, covering 7 emotion categories. The development of the MELT dataset mainly involves dialogue segment screening, selection of the GPT-4o model, and prompt engineering design. This dataset aims to address the challenges of high annotation cost and poor consistency in speech emotion recognition (SER), and its annotation quality and the improvement in model performance have been verified via both subjective and objective experiments.




