最近特征线在音频分类中的应用  

Application of Nearest Feature Line in Audio Classification

作　　者：练芝飞[1] 徐荣聪[1] 

机构地区：[1]福州大学数学与计算机科学学院,福州350002

出　　处：《计算机工程》2011年第2期151-153,共3页Computer Engineering

摘　　要：通过提取基音频率、明亮度、带宽、过零率、响度、均方根、相邻点之间距离的均值和方差及Mel倒谱系数这8个特征构造特征集,在此基础上提出一种基于最近特征线的音频分类算法,对其进行枪声、鞭炮声、喇叭声及说话声的分类实验中,结果表明,该算法的分类效果较好,错误率可低至11.76%。This paper constructs the feature set by extracting eight features including perceptual features like pitch frequency, brightness, bandwidth, zero-crossing rate, loudness, Root Mean Square（RMS）, the distance between the adjacent point of the mean value and Mel Frequency Cepstral Coefficients（MFCC）, and proposes an audio classification algorithm based on Nearest Feature Line（NFL）. It is applied to classification experiment with four audio including guns, banger, horn and talks, and the result shows that the algorithm is effective in classification and its error rate can reduce to 11.76%.

关 键 词：音频分类 最近特征线 音频特征选取 MEL倒谱系数 

分 类 号：TN912[电子电信—通信与信息系统]