Research And Implementation Of Music Retrieval System Based On Humming

With the rapid development of the Internet and software and hardware,the production and distribution of music have become more efficient and convenient,resulting in an explosive increase in the number of music.The rapid increase in the number of music has promoted the need for more and more efficient music retrieval methods.Humming-based music retrieval is a new type of music retrieval method,which belongs to content-based music retrieval.The humming audio of the humming retrieval system studied in this paper is in WAV format,and the songs in the database are in MIDI format.Humming melody extraction,humming position determination and melody matching are the three most important links in the humming retrieval process.This article focuses on the evaluation accuracy index MRR(Mean Reciprocal Rank),and is tested in the data set consisting of the IOACAS humming data set and 10,000 MIDI songs.In terms of humming melody extraction and melody matching,the main research work and innovations of this article are as follows:(1)In the humming melody extraction part,a method based on multiple probability distributions is proposed to calculate the observation probability in the PYIN(Probabilistic YIN)algorithm.This method uses different probability distributions to calculate the observation probability for audios with different periodic stability,which greatly reduces the probability that the PYIN fundamental frequency extraction algorithm recognizes the frequency doubling error as the fundamental frequency,and improves the robustness to noise,making the MRR index from 64.2% increased to 67.5%.(2)In the melody matching part,a melody matching algorithm based on the alignment of cutting points is proposed.The algorithm uses the note cut point alignment rule and uses the dynamic programming algorithm to solve the optimal alignment path,making full use of the information of each cut point,which is very robust to the miscut and miss cut of the note,and can be better Use the rhythm information of the melody and effectively improve the discrimination ability of the melody matching,so that the MRR index is further increased to 78.9%.(3)In the melody matching part,an improved DTW algorithm is first proposed.The improved DTW algorithm increases the constraint of the number of times of each point alignment,which can avoid excessive distortion without restricting the horizontal axis.Then a method for obtaining candidate melodies based on the improved DTW algorithm is proposed.This method uses the improved DTW algorithm to find the starting and ending points of the candidate melodies,which can more accurately find the candidate melodies and further increase the MRR index to 80.2%.