基于深度学习的音乐特征学习与分类

项目名称： 基于深度学习的音乐特征学习与分类

项目编号： No.61473167

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 张长水

作者单位： 清华大学

项目金额： 85万元

中文摘要： 音乐信息检索是一个有实际意义的研究领域。音乐分类是音乐检索的主要工作内容和技术手段。音乐分类的关键是音乐特征提取。音乐是一种错综复杂和结构细致的声音信号，目前音乐分类系统的性能难以满足实际需要。音乐特征高度抽象，而深层神经网络通过多层非线性操作，能够自动学习抽象特征，非常适合于音乐分类任务。本项目研究基于深度学习的音乐特征学习与分类，对于丰富深度学习的理论、扩大深度学习的应用范围、提高音乐分类的性能，都有重要的意义。课题具体内容包括：音乐数据库建立和音乐数据预处理；研究适合于音乐分析的新的深层网络结构；研究音乐深层网络结构的主动学习；研究多任务的深度学习网络结构；研究多任务的深度学习的作曲家分类与音乐情感分类；此外，还设计和实现一个用于算法实验和验证的基于深度学习的音乐分类实验系统。

中文关键词： 深度学习；深层神经网络；特征学习；音乐分类

英文摘要： Music Information Retrieval is a significant research area, which builds upon music classification as the main working area and key technology. Feature extraction is a crucial step for music classification. Music is basically a complex and delicate signal. Conventional music classification system cannot meet the practical need due to its highly abstract feature. While the deep neural networks can learn the feature automatically through multilayer nonlinear operation, and they are very suitable for music classification tasks. Based on music feature learning and classification of deep learning, the research will enrich the deep learning theory, improve the classification accuracy and expand its practical scope. Our research mainly focuses on: constructing the music datasets, pre-processing the music data, exploiting the novel deep neural networks for music analysis, studying active learning and multitask learning for deep networks, and proposing multitask deep learning algorithm and its application to composer and emotion classification. In addition, we will design and implement a music classification system based on deep learning for experimental evaluation.

英文关键词： Deep Learning;Deep Neural Networks;Feature Learning;Music Classification