Research On Music Generation Models Based On Heart Sound Features

Heart disease is a major health hazard in today’s society,with 17.7 million people dying each year from heart-related diseases.Heart sound classification is of great research importance as the main tool for heart disease pre-screening,which requires a medical professional to listen to the heart sounds and determine whether they contain pathological murmurs for diagnosis.However,it is difficult for the general public to directly classify raw heart sounds due to the expertise required.In order to lower the threshold for self-examination of heart sounds by the general public,this project investigates the method of converting complex heart sounds into more easily recognizable and more differentiated music files through deep learning methods,which enables more complex heart sound classification work through simpler music classification.Therefore,the research work in this thesis is divided into the following points:The purpose of studying heart sound music generation method based on heart sound Mel-spectrogram is to convert heart sounds into pictures first,and then realize heart sound music generation by using the more popular deep learning related methods nowadays.In this thesis,the heart sounds and music are converted into Mel spectrograms and the dataset is constructed,and the improved Cycle GAN is used for style conversion to realize the conversion from heart sound Mel spectrograms to music Mel spectrograms.After obtaining the heart sound music Mel spectrogram,a picturebased heart sound music generation scheme is realized by generating heart sound music using Wave Net vocoder,while the authenticity of the generated music and its ability to reflect the original heart sound type(which can be specifically expressed as the accuracy of classifying the original heart sound based on heart sound music)are compared by subjective research.The purpose of this study is to generate music based on symbolic music generation by constructing relationships between notes and chords and heart tone features.In this thesis,this thesis firstly extracted features based on Mel-scale Frequency Cepstral Coefficients(MFCC),and then extracted notes and chords from the music data set and encoded them.The training set is constructed by aligning the Mel-scale Frequency Cepstral Coefficients with the note encoding,and the note encoding is obtained by using the HBSM-NET model,and then the music21 library is used to arrange the music.A comparison is made between the symbolic and pictorial music generation methods.In order to put the target algorithm into daily application,the research and development of heart sound music generation system is carried out based on the heart sound music generation algorithm proposed in this thesis.The system mainly realizes the functions of saving the user’s heart sound,heart sound music generation and online display,heart sound music generation record keeping,etc.The main goal is to provide users with an interactive platform for generating heartbeat music.