基于最小距离的音频分类方法的研究

基于内容的音频分类是一个有趣并有重要意义的问题。音频分类技术包括音频特征抽取和分类器两个基本部分。如今，基于内容的音频自动分类技术已经有了很大的发展。然而，现有的基于内容的音频自动分类方法在分类的准确性、有效性和算法复杂度等诸多方面存在一定的不足，探索性能更佳的方法就成为了该领域的研究热点。提取了基于内容的音频分类所使用的音频特征，得到了基于帧的音频特征和基于片段的音频特征两个层次的特征，并提出了一种基于MFCC的简化的特征；选取了最小距离分类器中的最近邻分类器和K近邻分类器，对这几种典型的音频分类器进行研究，进行仿真实验，分析了实验结果；最后设计并仿真了经过改进的最小距离音频分类器，它的性能相对于原有的最近邻和K近邻分类器有一定的提高，并具有很低的系统复杂度和很短的分类时间。

esearch on Audio Classification Methods Based on Minimum-distance

Content- based audio automatic classification is an important and significant problem. The audio automatic classification techniques consist of two basic sectors which are the obtaining of the audio features and the classifying. By now, content - based audio automatic classification techniques develop much. But the present content - based audio auto- matic classification methods have some defects on the veracity, efficiency and algorithm complexity. Thus there is a trend to find a better method. Two levels of features used in content - based audio automatic classification are obtained, which are audio features based on frames and audio features based on clips. A simplified feature based on MFCC is proposed. The nearest neighbor and K - nearest neighbor classifiers are researched, the simulation result are analyzed. Finally, an improved minimum - distance classifier is proposed and simulated. It has a good performance relative to these of the nearest neighbor and K - nearest neighbor classifiers. And it has very low system complexity and very few classifying time.