摘要
基于相空间重构理论和主成分分析理论,对不同乐器产生的音频时间序列进行处理。通过分析各类乐器的高维特性,采用概率密度函数来刻画各个乐器在相空间中的差异,然后将概率密度函数的参数与其他音色特征相结合,采用柔性神经树作为分类器,提出一种新的乐器分类方法。柔性神经树能够解决人工神经网络结构的高度依赖性问题,还具有较高的识别率。实验表明,该分类器与BP神经网络和支持向量机比较具有较高的平均分类准确率和较低的均方根误差值。
Based on phase space reconstruction theory and principal component analysis theory, we process audio time sequences produced by different musical instruments. A musical instruments classification method is proposed by analysing the high dimensional properties of various instruments and adopting probability density function to describe the difference of each instrument in phase space, and then combining the parameters of probability density function with other timbre features as well as using flexible neural trees （FNT） as the classifier. The FNT can solve highly structure dependent problem of the artificial neural network and also has higher recognition rate. Experiment indicates that compared with BP neural network and support vector machine, this classifier has higher average accuracy rate of classification and lower RMSE value.
出处
《计算机应用与软件》
CSCD
2015年第2期159-162,214,共5页
Computer Applications and Software
基金
国家青年科学基金项目(61301250)
山西省人力资源与社会保障厅山西省留学人员科技活动项目(20121030)
山西省科技厅山西省国际科技合作计划项目(201281036)
山西省研究生优秀创新项目(20123107)
关键词
乐器分类
相空间重构
主成分分析
概率密度函数
柔性神经树
Musical instrument classification
Phase space reconstruction
Principal component analysis
Probability density function
Flexible neural tree
