乐器识别中的时序特征整合  

Temporal feature integration for automatic instrument identification

作　　者：范丽[1] 许洁萍[1] 王君[1] 

机构地区：[1]中国人民大学信息学院,北京100872

出　　处：《清华大学学报（自然科学版）》2009年第S1期1288-1293,共6页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(60873098)

摘　　要：特征整合是指在一个长时窗内将短时窗内的特征向量整合为一个新的单独特征向量的过程。音色是一种长时时序特征,目前常使用的均值和方差获得的长时特征,很难保持音色的时序特性。该文在多变量自动回归模型(multivariate auto regressive,MAR)特征的基础上,利用音乐最小单位音符为处理单元,提出了动态多变量自动回归模型(dynamic MAR,DMAR)特征,实现了音色时序特征整合;并将该特征向量应用于乐器音色识别中,同时在乐器识别的后处理中,利用加权平均求概率的方法去判断歌曲所属类别,使8种乐器的平均识别率从75.7%增加到87%,取得了较大提高。Temporal feature integration is the combining all the short-term features into a single long-term feature vector to capture the relevant temporal information in musical signals over a long time.Timbre is a long-term temporal feature that is difficult to identify,so the average and variance of short-term features are often used as long-term features but they fail to capture the temporal correlations among the individual feature dimensions.This paper presents a dynamic multivariate autoregressive feature model(DMAR) based on a multivariate autoregressive feature model(MAR) to solve this problem.A note onset detection algorithm is used to get a feature vector for each onset instead of a constant faster feature extraction rate.This technique can be applied for automatic instrument identification.Songs can also be categorized based on a weighted average probability method.The instrument identification rate reaches 87% indicating that this method gives good automatic instrument identification.

关 键 词：乐器识别 时序特征整合 特征向量 支持向量机(SVM) 多变量自动回归模型 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]