Abstract: Among all kinds of music styles, Beijing opera is very familiar in China. In this paper, we present a study on content-based classification among five kinds of typical aria and accompaniment of Beijing opera (including Sheng, Dan, Jing, Wenchang, and Wuchang), using audio analysis techniques together with pattern recognition techniques. A comparative evaluation between seven different classifiers is carried out on a testing database of 266 segments, and the results show that the BP neural network classifier (BPNNC) works best, and its average classification accuracy can achieve 88.7% in the five-class classification problem.

Key words: Beijing opera, audio analysis, pattern recognition

摘要： 本文使用音频分析技术和模式识别技术相结合的方法对传统京剧中的3种典型角色（生、旦、净）的唱腔和2种典型的纯伴奏形式（文场和武场）进行了基于内容的自动分类研究。我们的实验测试数据包括266个片段，来自于许多著名京剧演员如梅兰芳、袁世海、于魁智等人的演出。实验结果表明，对于5类分类问题可以达到88.7％的平均分类正确率，对于有伴奏下的唱腔和纯伴奏之间的2类分类问题可以取得高达96.6％的平均分类正确率。这一结果对京剧的进一步研究有着重要意义。

关键词: 京剧, 音频分析, 模式识别