Research On The Key Technology Of The Content-based Music Retrieval

Music is an important kind of audio data. The traditional music retrieval system is based on the key words, and the use is limited. With the rapid development of the Internet and the popularization of digital devices, rapid and effective searching for desired audio data becomes an important and challenging research topic. In this dissertation, lots of exploratory research work has been done around some key techniques of Content Based Music Retrieva—CBMR, which include feature extraction, similarity measure method, retrieval algorithm, end point detection, continuing speech identification and so on. The research fruit has got application in the musical retrieval system. This research work is under the support of Natural Science Foundation project of China "research on the key technology of audio recognition and retrieval based on the web (No. 60673100)". The main contributions of this dissertation are summarized as follows:1) The processing frame and the symbolic system of the music retrieval system are given. The research solution is given based on the problem existing in the music retrieval. The music Gepu mapping methods and the symbolic system are defined, which is the foundation of the content-based music retrieval.2) The algorithm studying on the music features. Music features can be divided into two parts, i.e., low level feature (melody contour) and high level feature (music score).a) A feature extraction algorithm of the music rhythm contour is proposed according to the lower features, which can accurately extract melody contour features from the humming voice. However, this method requires user to grasp the pitch accurately.b) The technology of the speech recognition is used in the music retrieval system according to the high level feature. The End-Point Detection method is used to detect the start-point and end-point of the note in the continuous speech signal composed of music score. Based on time series analyzing method-PAA and entropy an end-point detection algorithm—PAAEPD(Piecewise Aggregate Approximation End Point Detection) is proposed. Then the goals music can be got by humming the music score. This approach requires user pronounce accurately.3) The research of music retrieval algorithm.a) According to the music feature-time series, the character distance and string distance are defined, and a calculating method of string similarity is proposed. The necessary of retrieving approximate results over complex data such as audio is analyzed, which is irregular and not compared, as well as the limitations of the traditional approximate string matching algorithm and time series similarity search algorithm. Then based on them the method is proposed. At last, it has been used in the music retrieval system and the top-10 retrieval rate is 92% to the user who can grasp pitch accuratly.b) According to the music feature-sentence unit, a retrieval method based on sentence is proposed. The Rhythm Fluctuate and length of each sentence are fixed. So they can be used to reduce searching range.c) According to the singer's habit, a multi-sentence retrieval method is proposed. And the candidate music segment should be built in the search time. The method has been given in the paper.4) The studying of the music database.The original Gepu data is the source of the database. The feature in the music database, as well as their meaning and calculation method is given in the paper.5) A content-based music retrieval system is realized based on the algorithms studied in this paper, and retrieval results are good. The study results can be used in other similar fields.