Research On Melody Feature Representation For Query By Humming

Query by Humming（QBH）is a research hotspot in the field of music information retrieval,which can provide users with a convenient and intelligent retrieval experience.The QBH needs to retrieve the song melody characteristics that match the humming melody characteristics,so as to determine the category of the humming song according to the degree of matching of the melody characteristics.Therefore,how to condense and match representative feature representations from audio data is the most noteworthy issue in QBH.For this reason,it is necessary to explore the methods that can express the characteristics of melody and effectively express the characteristics of songs with large differences and different versions.Based on this motivation,this work will conduct research from the perspective of melody feature representation.The specific research contents and innovations are as follows:（1）Aiming at the problem of unstable melody features,a QBH method based on melody feature clustering and optimization is proposed.This method introduces the optimized initial clustering center k-means in the QBH system.By clustering melody features,the structural similarity between different melody features could be learned fully,the melody features with similar structures are divided into the same cluster and the clusters are numbered,and effective labels are obtained.The clustered labels are used as the labels of the convolutional neural network and feature extraction is performed to obtain a more discriminative melody feature representation.Thus,the robustness of the melody feature is effectively improved.After obtaining the melody feature representation,it is necessary to search the cluster category as the matchinglabel.Experimental results show that the proposed method can effectively improve the performance of the humming retrieval system.（2）Aiming at the problem that the deep feature relationship between melodies cannot be captured accurately,a melody feature representation method based on deep metric learning is proposed.This method uses the deep metric learning optimization function to measure the similarity of melody features,and maximizes the similarity relationship between melody features,so that the features of different humming versions of the same melody are more compact and the features of different melody are more scattered in the spatial distribution.Specifically,this method maximizes the similarity between the same melody features and minimizes the similarity between the different melody features as the learning goal of the neural network.Then the residual network is used to extract the melody feature,and then the gated recurrent unit is used to capture the time information of the melody feature,and finally the attention mechanism is used to pay attention to the importance of the melody feature.Using this network and the optimization goal could obtain a feature representation with timing dependence,so as to describe the melody feature more accurately.Experimental results show that the proposed method can greatly improve the accuracy of the humming retrieval system.