Humming Music Retrieval Based On Melody Feature Extraction

With the development of multimedia technology and network technology, and the rapid growth of the number of multimedia information, in order to take full advantage of the existing multi-media resources, it is necessary to do research in multi-media retrieval techonology. How to get the suitable music information from a large number of audio information is the target of hamming retrieval technology. Also, it is one of the main method in multi-media retrieval.The purpose of this research project is through analyzing the user's humming input, extracting the note sequece, and then get the right name of the song back to the user. This is the target of content-based music retrieval. In this paper, our research mainly in the following aspects.1) Pitch is the basis for identify notes. Firstly, we analysis some of the existing pitch extraction alogrithm. Taking the real-time nature of music retrieval, we determine our identification using YIN algorithm and CUEX algorithm, also we make some improvement based on them.2) Do some research in melody extraction of MID files. Proposed melody extraciont method based on track statistic features, using improved artificial neural network.3) We design and implement a humming content-based music retrieval system. After downloaded the MID files, we can get the melody database using the melody extraction algorithm. We use the Chinese word segmentation approach based on melody database to create a music melodic contour database. Then, based on melodic contour database and melody word database to establish inverted index, then the text retrieval methods can be applied to the music retrieval.During note identification and MIDI melody extraction, we have got a higher accuracy rate, and on this basis, we can get a music thesaurus and indexing system which can reflect the melody characteristics of songs. As the music lexicon is a very new concept, the retrieval accuracy can reach about 46%. Some conclusions have been made by experiment results. These conclusions can reflect some of the features for humming music retrieval. This can provide a theoretical basis for further research.The methods proposed in this thesis can be used to other speech research. The study of humming method can be quickly applied to other areas which presents the good scalability of this system. In addition, the algorithms can bring a certain reference value for the retrieval and classifcation technology.