基于中高层特征的音乐情感识别模型

发布时间：2018-12-14 09:20

【摘要】：为提升音乐情感识别的准确率,提出基于中高层特征的音乐情感识别模型,摒弃频谱特性、色度、谐波系数等低层特征,以更接近于人认知的中高层特征包括和弦、节拍、速度、调式、乐器种类、织体、旋律走势等作为情感识别模型的输入。建立一个包含385个音乐片断的数据集,将音乐情感识别抽象为一个回归问题,采用机器学习算法进行学习,预测音乐片段的8维情感向量。实验结果表明,相比低层特征,采用中高层特征作为输入时的准确率R2能够从59.6%提高至69.8%。
[Abstract]:In order to improve the accuracy of music emotion recognition, a music emotion recognition model based on middle and high level features is proposed. The lower features, such as spectrum characteristics, chroma, harmonic coefficients and so on, are abandoned, so that the middle and high level features, which are closer to human cognition, include chords and rhythms. Speed, mode, musical instrument type, texture, melody trend, etc., as the input of emotion recognition model. A data set consisting of 385 pieces of music was established. The recognition of music emotion was abstracted into a regression problem. Machine learning algorithm was used to study and predict the 8-dimensional emotion vector of music segment. The experimental results show that, compared with the lower level features, R2 can improve the accuracy from 59.6% to 69.8% when using middle and high level features as input.
【作者单位】： 复旦大学电子工程系;复旦大学信息学院智慧网络与系统研究中心;
【分类号】：TP181;TN912.34
，

本文编号：2378372