Research And Implementation Of The Content-Based MP3 Music Retrieval

Content-based music retrieval uses the content of music itself as query, such as rhythm, melody, chord, accompaniment, timbre etc., to get the music resource which meets the query condition or contains same or similar music clip in a music database. The audio media in the music database may be of different formats, such as MIDI, MP3 and other wave audio. These queries can be inputted into the retrieval system using humming or existing score input tools. This dissertation researches some technology of content-based music retrieval and implements a flexible prototype system that has a variety of query input methods.At present, most work in music retrieval has mainly focused on the MIDI type retrieval, and little attentions are put on the MP3 type. Also, most of these systems provide only one query input method to the users. In view of different users may be with different music literacy and have different music resources, the technology of multiple input modes in one music retrieval system is proposed in this dissertation. Our system provide stave input mode, approximate numbered musical notation input mode, humming input mode, monophonic MP3 file input mode and polyphonic MP3 or other wave file input mode. Currently the technology researching on the pitch recognition of the wave music is not enough mature, and the rhythm and melody of the user's humming query is not able to completely accurate. These factors all makes the final melody characteristics extracted from the query will be different from the expectations of user. Of course, this will affect the accuracy of retrieval. Therefore, our system also allows user to judge and correct the melody characteristic extracted by the system according to their own music ability through a variety of interactive tools, which not only enhance the accuracy of retrieval, but also improve the availability of the system.In the melody match module, we use different melody feature matching method for monophonic music and polyphonic music according to their different characteristics. For the former, this dissertation presents a new secondary warp matching algorithm, which includes the rest merging and long vowels segmentation. And we proved effectiveness of the proposed algorithm through analyzing the result of experiment. For the latter, we introduce the harmony model proposed by Jeremy (2004) to our system, which was applied to retrieval harmonic MIDI music database using harmonic MIDI query. We use this technology to retrieval MP3 music database by harmonic MP3 music files, and propose two retrieval models called harmony model based on HMM and harmony model based on Markov model respectively. The experimental results indicate that the two models all have certain practical value.