In this paper we proposed a music matching method, called the RLCS (rough longest common subsequence) method. It is an improved version of the LCS to avoid some problems occurring in global alignment matching. First a rough equality for two notes is defined for constructing the RLCS of two music fragments. The length of the RLCS of two music sequences defined in this work is a real number, called a weighted length. It is evaluated according to the degree of similarity of every pair of matched notes from the two sequences. This method takes into account both the width-across-query (WAQ) and the width-across-reference (WAR) and combines them with the weighted length of the corresponding RLCS to define a score measurement for the RLCS. The measurement associated with WAQ and WAR enables the proposed method to tolerate dense errors. A dynamic programming algorithm is presented for simultaneously calculating the weighted length of RLCS, the WAQ, the WAR, and the score to determine the RLCS. As a result, the proposed method can perform the matching in a better and simpler manner. In order to speed up the matching process, we use the filtering algorithm proposed by Tarhio and Ukkonen [22] to filter the reference and discard most of the reference areas that do not match. We applied the proposed algorithm to content-based music retrieval. The experimental results showed that with our proposed algorithm the retrieval system provides a higher retrieval rate than that with the local alignment method proposed by Suyoto et al. [20]. The use of the filtering algorithm has been shown to greatly reduce the computation time for exact matching and for approximate matching with a low error tolerance.