"Memo2496: Expert-Annotated Dataset and Dual-View Adaptive Framework for Music Emotion Recognition"

"Music emotion recognition delineates and categorises the spectrum of emotions expressed within musical compositions by conducting a comprehensive analysis of fundamental attributes, including melody, rhythm, and timbre. This task is pivotal for the tailoring of music recommendations, the enhancement of music production, the facilitation of psychotherapeutic interventions, and the execution of market analyses, among other applications. The cornerstone is the establishment of a music emotion recognition dataset annotated with reliable emotional labels, furnishing machine learning algorithms with essential training and validation tools, thereby underpinning the precision and dependability of emotion detection. The Music Emotion Dataset with 2496 Songs (Memo2496) dataset, comprising 2496 instrumental musical pieces annotated with valence-arousal (VA) labels and acoustic features, is introduced to advance music emotion recognition and affective computing. The dataset is meticulously annotated by 30 music experts proficient in music theory and devoid of cognitive impairments, ensuring an unbiased perspective. The annotation methodology and experimental paradigm are grounded in previously validated studies, guaranteeing the integrity and high calibre of the data annotations."

音乐情绪识别（Music emotion recognition）通过对旋律、节奏、音色等基础属性开展全面分析，对音乐作品中传递的情绪光谱进行界定与分类。该任务对于定制化音乐推荐、优化音乐制作、辅助心理治疗干预以及开展市场分析等诸多应用场景均具有关键意义。其核心基石在于构建标注有可靠情绪标签的音乐情绪识别数据集，为机器学习算法提供必要的训练与验证工具，从而保障情绪检测的精准性与可靠性。本文提出了包含2496首器乐作品的音乐情绪数据集Memo2496（Music Emotion Dataset with 2496 Songs），该数据集标注了效价-唤醒（valence-arousal, VA）标签与声学特征，旨在推动音乐情绪识别与情感计算领域的发展。该数据集由30名精通音乐理论且无认知障碍的音乐专家进行精细化标注，确保了视角的客观性。其标注方法与实验范式均基于已验证的既往研究，保障了数据标注的完整性与高质量水准。