基于音频特征的乐器分类研究

The study of musical instrument classification based on audio features

针对现有乐器分类研究中存在的使用特征量过多、分类准确率有待提高等问题,提出了一种特征量少、准确度高的乐器分类方法。基于Relief算法的主成分特征提取方法,计算出各特征量的权重,设计3层的神经网络分类器。根据所提算法和分类器,使用8项音频特征与传统的24项MFCC特征,分别对中西方9种乐器进行了分类实验,并分别使用权重最高的4、5、6项特征进行分类实验。结果表明,所提出的音频特征相比于传统MFCC特征对乐器分类的平均准确率更高,达到94.84%,且特征量更少,说明基于Relief算法的主成分特征提取方法能有效减小低相关性特征对分类准确率的影响。

To solve the problems in musical instrument classification studies like using too many features,lowclassification accuracy,etc,we proposed a musical instrument classification method with less features and high accuracy.We calculate the weight of each characteristics using the principal component feature extraction method based on Relief algorithm,and design three-layer neural network classifier.According to the proposed algorithm and classifier,we conduct the classification experiment between eight features proposed in this paper and traditionally twenty-four MFCC features,which include nine musical instruments consist of Chinese and western musical instrument,and accomplish the classification experiment by using the fourth,fifth and sixth characteristics with highest weight respectively.Results show that features proposed in this paper is fewer than MFCC features,and can get higher average accuracy which reached 94.84%.We can draw a conclusion that the principal component feature extraction method based on Relief algorithm can reduce the influence of low correlation characteristics on classification accuracy effectively.