一个基于相关反馈的神经网络音乐情感分类器构建算法

A music emotion classifier construction algorithm of neural network based on relevant feedback

目的构造音乐情感分类器,为计算机理解情感提供途径。方法首先分析现有音乐情感模型,然后提取音乐情感特征并采用神经网络构造音乐情感分类器,最后采用相关反馈对分类结果进行标注并进一步训练分类器,直至训练样本的错误分类率在一定误差范围内。结果对测试样本进行情感分类,达到了最高89%的分类准确率,实现了音乐情感分类器的构建和音乐片断的情感标注。结论研究成果是音乐情感检索的基础工作,也可用于其他音频的情感识别和分类。

Aim The emotion is the essence feature of music. The construction of music emotion classifiers provides the solution by which computer understands the emotion. Methods This paper begins with the analysis of the existing music emotion models. Then the music emotion characteristic is extracted, and the music emotion classifiers are constructed based on neural network, which is trained based on the right records that labeled in the database by relevant feedback, until the error rate of training sample classification is within a certain range. Results The classification accuracy amounts to 89% to the test samples. The classifiers are constructed and the music clips are labeled. Conclusion The results can be used to other audio emotion recognition and classification area, and it is the foundation of music emotion retrieval.