Scholarlink

<p>Music is a key form of artistic communication, closely connected to the perception, evocation, and expression of emotions. Thus, the development of prediction models that estimate perceived emotions in music must consider changes in music audio waves, as this is highly correlated with perceived emotional transitions. Within the music emotion variation detection field, transformer-based models have been gaining attention due to their ability to capture long-range dependencies. This work uses a dimensional approach based on Russell’s Circumplex model of Affect to estimate values in the valence/arousal plane, a curve introduced as the <i>emotional trajectory</i>. Based on the available dynamic annotations of the MediaEval Database of Emotional Analysis of Music, an adapted transformer model architecture is implemented to predict emotional transitions directly from music audio features. Our model achieves state-of-the-art results, with RMSE values of 0.217 and 0.261 for arousal and valence dimensions, respectively, obtaining an approximation of the expected emotional trajectory evoked by a song. Furthermore, we applied a windowed concordance correlation coefficient-based metric to capture trend variations in the estimated trajectory, demonstrating promising results in recognizing emotional transitions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A transformer approach for predicting transitions in emotional trajectory directly from music audio

  • 期刊论文

  • 发表日期：

    2025年6月2日

查看全文
  • Pascal A. Stoller,
  • Mauricio Araya

摘要

Music is a key form of artistic communication, closely connected to the perception, evocation, and expression of emotions. Thus, the development of prediction models that estimate perceived emotions in music must consider changes in music audio waves, as this is highly correlated with perceived emotional transitions. Within the music emotion variation detection field, transformer-based models have been gaining attention due to their ability to capture long-range dependencies. This work uses a dimensional approach based on Russell’s Circumplex model of Affect to estimate values in the valence/arousal plane, a curve introduced as the emotional trajectory. Based on the available dynamic annotations of the MediaEval Database of Emotional Analysis of Music, an adapted transformer model architecture is implemented to predict emotional transitions directly from music audio features. Our model achieves state-of-the-art results, with RMSE values of 0.217 and 0.261 for arousal and valence dimensions, respectively, obtaining an approximation of the expected emotional trajectory evoked by a song. Furthermore, we applied a windowed concordance correlation coefficient-based metric to capture trend variations in the estimated trajectory, demonstrating promising results in recognizing emotional transitions.