Research On Music Information Retrieval Technology Based On Content And Semantic

As the size of audio and music collection dramatically increase in internet,Content-based Music Information Retrieval(MIR)is gaining widespread attention and can be very helpful,since it forsakes the need of keyword.Often,it consists of a form of query-by-example,such as singing,humming or playing a sample of the piece,as a query to the database.Most of the music retrieval is based on text retrieval as the main means,requiring a lot of manual tagging,which greatly hindered the retrieval and dissemination of music.In addition,this becomes the bottleneck problem of the development of digital music industry.Music is the product of human thinking,with physical waves as the carrier,conveying people's understanding of life and subjective feelings.But there is a so-called semantic gap between the low-level musical features and the semantic features of music".In this paper,the signal processing and music analysis,music content as a means of expression,semantic description of music as the carrier,and gradually realize the more natural,close to the people of music perception to music database retrieval,which conform to the user's subjective experience of music files.Query by humming(QBH)refers to music information retrieval systems where short audio clips of singing or humming act as queries.Melody is considered as the most important feature in the queries and the songs.This paper proposes a QBH system using melody matching model based on the genetic algorithm and improving the result by a learning to rank approach.A template of the query music is constructed by a dynamic threshold note segmentation method,and a melody contour aligning algorithm based on GA is proposed,which is used to align and correct the input pitch template.The proposed learning to rank algorithm based on Support Vector Machines is used to improve the rank result.The validity of the algorithm is presented by the prototype of QBH system and effects of the algorithm are also shown by the experiment result.An audio information retrieval model based on Manifold Ranking(MR)and improving ranking results by relevance feedback algorithm is proposed in this paper.Timbre component has been employed as the main feature.To compute the timbre similarity,it is necessary to extract the spectrum features for each frame.The large set of frames is clustered by a Gaussian Mixture Model(GMM)and Expectation Maximization.The typical spectra frames from GMM is drawn as the data points,manifold ranking assigns each data point a relative ranking score,which is treated as a distance instead of traditional similarity metrics based on pair-wise distance.Furthermore,manifold ranking algorithm can be easily generalized by adding these positive or negative examples by relevance feedback algorithm,and improves the result.Experimental results show the proposed approach is effective to improve the ranking capability of the existing distance functions.Query by semantic description(QBSD)is an increasingly prominent technology that allows users to search music through semanticallyannotated song database.A QBSD system can be defined as a supervised Multi-class labeling(SML),by which a song could be mapped to semantic space.In this paper,we present a method for querying by semantic description in content based music retrieval;a song can be used as query and transformed into semantic vector by a convolutional neural network.The experiment demonstrated that our proposed method can catch results relevant music accurately and efficiently in semantic space.