An Evaluation of Feature Extraction for Query-by-Content Audio Information Retrieval

Content-based audio information retrieval is one of the most interesting and fast-growing research areas. Suitable feature sets can help to reduce the tedious computation time and speed up retrieval. In this paper we report a study of the music spectral properties aimed at the acoustic-based music data similarity measurement and show that the spectral features of adjacent frames are highly correlated. Based on such a case study we mainly focus on making an evaluation of feature choice in the three aspects: storage, computation and retrieval ratio. The extensive evaluations confirm the effectiveness of feature merge in quickening sequence matching for query-by-content audio retrieval and show that MFCC with feature merge is the best tradeoff among storage requirement, computation cost and retrieval ratio.