HeartBeats:The Research Of Music Generation Model

Warm-up is an important part of exercise,and it not only makes one to fit vigorous exercise gradually but also can reduce the risk of occurrence of danger.Relevant studies have shown that when the warm-up heart rate reaches 50%to 60%of the maximum heart rate,the exercise state of the sportsman is close to optimal.Strong music has a positive impact on physical and mental states by stimulating the cerebral cortex,thus,music has become one of the appropriate stimuli to adjust the heart rate in exercise.Percussion music is a type of music which is mainly played by knocking,striking,shaking and rubbing instruments.The keen sense of rhythm is a common feature of percussion music and exercise.Therefore,Percussion music is one of the appropriate stimuli in exercise.At present,the research of exercise heart rate adjustment by music is mainly confined to the field of sports science.The subject of research is generally the professional athlete,and the choice of music for the adjustment depends on subjective judgments.Therefore,there is uncertainty that the method of subjective choice for music cannot adjust the heart rate adaptively.At the same time,the research object is often a professional athlete,which has limitation.A music generation model that targets the heart rate adjustment of non-athletes is proposed in this paper,which is called HeartBeats.In the part of feature processing,we chose the drum kit to play the music which is the most representative percussion instrument at first.Secondly,we collected 50 songs in MIDI format.Thirdly,we extracted the drum tracks from these songs to build the database.At last,we converted the data into features for processing.In the part of model training,we used the reinforcement learning technology to build the music generation model,and add the generative adversarial mechanism by training a discrimination model for predicting the reality of music as the feedback of reinforcement learning.Finally,a system application was developed and 30 participants were recruited to use our application.We collected the heart rate data of the subjects in the experiment,and we did statistical analysis of the results to verify the proposed model.The results and innovations of this paper mainly include:(1)We build a music generation model for exercise heart rate adjustment.We use the reinforcement learning to solve problem for functional music generation adds a discrimination model for predicting the reality of music to set up the generative adversarial mechanism.The correspondences between three elements of the reinforcement learning are that:1)music context and heart rate context represent the state 2)the way of the next beating reprensents the action 3)the heart rate change and the reality outputted by discriminator represent the feedback.(2)We conducted an empirical study of the music generation model proposed in this paper,recruiting 30 subjects for experiments.The results of experiments show that:1)When subjects sitted quietly and listened to the music composed by HeartBeats,the average heart rate was increased by 14.52%and the peak heart rate was increased by 16.64%compared with the resting heart rate.2)When subjects walked slowly and listened to the music composed by HeartBeats,the average heart rate was increased by 26.82%,and the peak heart rate was increased by 35.51%compared with the resting heart rate.3)When subjects walked slowly and listened to the music composed by HeartBeats than not listening to music,the heart rate is effectively closer to the warm-up heart rate range(50%to 60%of maximum heart rate).This paper is a cross-study of deep learning model improvement,empirical experiment design and human-computer interaction system.The purpose of this study is to promote the active regulation of exercise warm-up heart rate.The experimental results show that HeartBeats can effectively adjust the heart rate to the best warm-up range,and to a certain extent,HeartBeats solves the uncertainties and limitations of randomly screened music.This article provides exploratory applications and empirical results for the field of music generation,which is of great value in the further development and perfection of the interactive model of "human,music and artificial intelligence".