Abstract: Audio datum bears the characteristics of huge amount of data and high dimensionlity，directly indexing the raw data readily causes the curse of dimensionlity.In this paper，a new approach of audio data reduction based on fuzzy rough sets is proposed，which fuzzily discretizes the audio depending on affiliation，reduces the redundant data by the knowledge expressible ability，evaluates the knowledge core according to the equal classification ability and discernibility.Experimental evaluations show that the proposed algorithms can produce audio data faith to the original data and improve the efficiency of audio indexing.

Key words: audio indexing, fuzzy rough set model, audio feature, feature reduction

摘要： 音频具有数据量大、维数高等特点，直接进行音频检索会造成“特征维数灾难”，因此有必要从音频提取最能表现音频特征的音频帧。提出一种基于模糊粗糙集模型（Fuzzy Rough Set Model，FRSM）的音频数据约简算法，根据隶属度对音频数据进行模糊离散，基于知识表达能力约简属性，以等价划分计算具有等同分类能力的知识核。实验结果表明，该算法能够得到最小约简，并且最大程度地保持音频特征，提高检索效率。

关键词: 音频检索, 模糊粗糙集, 音频特征, 特征约简