基于人类发声系统复杂的时变的非线性本质，研究基于非线性动力学的声信号递归特性描述和分析方法，在此基础上构造歌曲的相似性度量算法，并探索其在快速翻唱歌曲检索系统中的应用。1）拟引入人类听觉可计算模型，研究结合乐理和听觉感知模型的高效的特征提取算法；2）拟采用计算机视觉分析的方法对音乐耳蜗图进行分析并提出可对翻唱版本和非法拷贝进行区分的音乐指纹算法；3）拟研究基于声信号动力学特性递归分析的音乐相似性度量算法；4）最后拟提出基于图形处理器的检索系统的并行实现方案。本课题预期的研究成果可为数字音乐信号处理研究提供新的思路、新的分析工具和新的特征提取手段。它不仅可用于解决大规模翻唱歌曲检索的应用难题，还可广泛应用于数字音乐版权管理和海量音乐数据的检索与管理。

Due to the complex and time-varying nonlinear nature of human vocal system, we study the description and analysis methods of song's recurrence characteristics based on nonlinear dynamics. On this basis, we construct a quantitative similarity measurement of songs and explore its application in a fast cover song retrieval system. We intend to 1）introduce the computable auditory model and study efficient feature extraction algorithm combining music theory and auditory model; 2）to apply computer vision analysis to the music's cochleagram and then propose fingerprinting algorithm that can be used to distinguish between the cover versions and illegal copies; 3）to study the music similarity measurement algorithm based on recurrence analysis of dynamical characteristics of acoustic signal; 4）and finally to propose the parallel implementation scheme for retrieval system based on Graphics Processing Unit (GPU). The expected research achievements of this project provide new ideas, new analysis tools and new feature extraction means for digital song signal processing research. And it can not only be used to solve the application problems existing in large-scale cover song retrieval system, but also be widely used in musical rights' management and music collections's retrieving and organizing.