The Research Of Music Retrieval System Based On Humming

Search based content is very common in the field of querying by humming. It is used in speech recognition widely. The thesis aims to achieve a basic music retrieval system based humming. In order to complete the implementation of the system, we must solve the three major technical difficulties. The first issue we need to do is how to extract the music characteristic curve of the music melody. And the second issue we need to do is how to build the characteristic database of the music coded by midi format. And the third issue we need to do is how to search the target item from the music feature database, namely how to improve the melody match algorithm.The work on the first technical difficulty includes preprocessing in a music and the feature melody curve extracted from a preprocessed music. In order to get some clean music data, so we use Kalman filtering algorithm to get rid of the noise in the input music signal. And we use a first-order high-pass filter to pre-emphasis on the de-noised music signal. At last we use the Hamming window to frame the processed music. After this we can extract the music melody feature from the pre-processed music. Then we discussed a lot of existing feature extraction algorithms. Ultimately, we choose the improved average magnitude difference function method(AMDF) as the feature extraction algorithm. And we do some simulations on this algorithm using matlab tool. Finally, we complete the feature extraction from the input music.The work on the second technical difficulty is how to extract the feature curve from the midi files and how to build the music feature database using the extracted feature. In this chapter we do some analysis on the format of midi files and read all data in midi files by using the third-party library. After this we get the pitch feature curve of the midi files and write these data into files by improving the skyline algorithm and the optimal k track melody algorithm. The data in the file is the template feature supported the melody feature match algorithm.The work on the third technical difficulty is how to retrieve the desired items from the music feature database. We analyze the existing melody matching algorithms. These algorithms include string matching algorithm, dynamic time wrap(DTW) algorithm and then hidden markov model algorithm. Ultimately, we choose the dynamic time wrap algorithm as the feature matching algorithm. Taking into account the actual retrieval of the versatility, we make some improvements on the DTW algorithm. Wherever the user hums from the improved algorithm can recognize the humming melody fragments.At last, we use the above research results to build a simple music retrieval system based humming. We implement all functional modules mentioned and do a lot of work on this system. The results of the system can achieve the expected requirements.