Study And Implementation Of Pop Music Accompaniment Generation Algorithm Based On Deep Learning

Arrangement is an essential part of a complete pop music production process,and the traditional way of arranging accompaniment requires mastering the principles of instruments and takes a long time.To save time costs and lower the threshold of music creation,this thesis investigates the use of deep learning algorithms to generate music accompaniment to assist musicians in composing music.Moreover,this method can also be extended to other music types.However,in the current research of deep learning-based music generation,the effect of single-track generation and multi-track collaborative generation needs to be improved;in addition,the research on the generation of accompaniment with controllable attributes has been little discussed.In this thesis,we propose MuseFlow,a Flow-based music accompaniment generation model.The model uses practical pianoroll data for training,learns potential patterns in music,and arranges instrumental accompaniment such as drums,guitar,bass and strings for the main melody.In this thesis,comparisons are conducted on several music datasets,and the results show that MuseFlow can effectively improve the accompaniment quality and track harmony,providing a solution to generate high-quality and controllable music accompaniment.First,to represent music data appropriately,this thesis proposes a uint8 pianoroll method that can identify the starting point of notes,which can save storage space while preserving the original information of the music material.Second,in the process of algorithm research,this thesis proposes MuseFlow,a model for multi-track music accompaniment generation,to achieve a two-way mapping of music and Gaussian hidden variable space.The model inputs piano melodies and random variables to generate drum,guitar,bass and string accompaniments.Third,due to the memory limitation of the GPU,MuseFlow can only generate 8 bars at a time.In order to avoid the time limit,this thesis designs a sliding window generation method,which can generate long music accompaniment sequences.After that,we conducted some experiments to verify the effectiveness of the model.In comparison with other models,1000 test cases were randomly selected from LDP,FreeMidi and GPMD datasets to generate music accompaniment using MuseGAN,MMM and MuseFlow.The measurement results of in-track mass and track spacing show that the MuseFlow model produces better results in quality and harmony,and it is closer to real data in pitch and length distribution.The uint8 pianoroll method can further improve the note quality.In the experiments of attribute control,this thesis uses feature vector operations to achieve controllable generation of pitch and interval.Finally,based on the MuseFlow model,this thesis uses the Django framework to develop an online interactive automatic music accompaniment generation system,including music accompaniment generation,melody generation and score editing,for assisting musicians in music composition.