Humming-based Music Retrieval Technology

With rapid development of digital music, networking and mobile phone technology, linking the three to meet the needs of daily life, learning and entertainment has become the trend of the times. Traditional music retrieval based on text must require users to give keywords, such as music name, the author, the lyrics and so on. This approach has significant limitations. With the accelerating pace of life, people have been not satisfied with this text-based music retrieval. people might only remember the tune of a song, or perhaps a fragment of it. How to quickly and effectively retrieve music data by melody information has become a prominent problem. In this dissertation, we study and explore the music retrieval based on hummed melody.As a kind of content-based music retrieval, it doesn't need to mark the text, just by humming to search the music. For this purpose, we've done the following researches.(1) Introduced the research of query by humming, the basic knowledge of music and the basic method of signal processing.(2) In this dissertation, we used the pitch interval and the duration proportion as the music feature to avoid the pitch changing and uneven rhythm. In the stage of feature extraction, we first used ACF to estimate the fundamental period. Then, we compute the mean and standard deviation of short-term energy and set a dynamic threshold to realize note segmentation.(3) In the stage of melody matching, we designed an algorithm named Index-based Modified DTW. The new algorithm used pitch interval as the feature to avoid pitch correction, and set the boundary conditions to avoid the comparison from start to finish. At last, we add the rhythm information to compute the final similarity score. Taking into account the search efficiency, we used the method of N-Gram to build the index of the database. This dissertation proved the efficient of the algorithm, and the top-ten searching rate is 92%.