Research On A QBH System Based On Melody Information Invariance

Query by humming(QBH)as one of content-based retrieval method can provide users a way to find the target music in the music library by humming part of the melody.Because of the differences in the singer's personality,such as the different parts of the whole range,the partial pitch shift and the inconsistency of the humming rate and so on,the extraction of the stable humming features has been a research difficulty in the research of QBH system.For these three kinds of personality factors,how to maintain the invariance of melody information in the humming signal is the main research content in this paper.The thesis points out that the rate of humming rate is not unified as a breakthrough point.In order to improve the performance of the QBH system,the thesis deal with the typical personalized performance of human voice,and fully excavate the most essential stability information of humming melody information.The main research content and achievement are as follows:1.Improvement of pivotal points of humming melody segmentation and feature extraction in detail.This paper presents the feature extraction strategy that segmenting the melody first,and then extracting the characteristics of the segment deeply.Thus,it can not only improve the efficiency of humming characteristic index,but also excavate the essence of humming melody.The extreme notes and some midpoint notes of the melodic extreme notes are the most stable notes in melody information,which called the pivotal point of melody and the benchmark of melody segmentation.The thesis optimized the pivotal point extraction effect based on the relationship between notes in music.It can guarantee the segmentation accuracy in humming melody even under the condition of hum rate not consistent.Then extract the melody from the segments using he traditional feature extraction method including local unevenly distributed histogram statistics feature,the characteristics of the senses and the characteristics of rhythm and statistics.Besides,by the feature extraction of several continuous local statistics,the distinction between features is enhanced.Finally,the experiment results show that the method have good stability for the personality factors of humman voice.2.The thesis studies a method of humming feature extraction based on deep auto-encoder.Based on the correct segmentation of the melody and the powerful feature extraction ability of the sequence data by utilizing the neural network,the deep auto-encoder model is selected to get the coding features as the most representative of the melody in the segment for the final humming signal retrieve features.The deep-level feature that transforms the melody in the non-linear space many times and represents the most fundamental feature information of the melody.Experimental results show the effectiveness of the method.Finally,in order to inherit the advantages of traditional features and auto-encoder features,the traditional features and coding features are combined together to be the final humming retrieve features,and the stability of the joint feature is proved through experiments.