Research On Automatic Composition Technology Based On Deep Learning

In recent years,the rapid development of computer composition technology,the use of computer to create high-level music,is the common goal of all researchers.The mainstream direction of the current field is to simulate a certain kind of works through deep learning technology to achieve music creation.however,ignores song structure and generates music with a noticeable mechanical feel.The work is merely a combination of chords and unable to ensure melody sounds nice and sweet.In view of the issue that other automatic composition methods based on the whole song learning results in the lack of song structure,this paper,through the study of structural composition,proposes a Motif-subphrase-phrase model,starting from the composition mode of Motif-subphrase-phrase and using deep learning techniques to learn composition.The model simulates the way composers compose,which includes three steps to composing: Building subphrase from motif,phrase from subphrase and chords orchestration from phrase,adding more life to automation-generated songs and creating motif-based ones with distinct themes.In this paper,the research is carried out from the following three aspects.1.Based on the current situation that there is no motif data set,this paper proposes an approach to motif extraction called seq2 seq so as to make sure there are sufficient motif data sets for training provided for Motif-subphrase model.By manually labeling the current music data set with motif,this approach uses seq2 seq model to learn motif extraction to settle the problem of lacking motif in structural composition model.2.In order to solve the problem that the previous automatic composition methods used the whole piece as the training set,resulting in the lack of song structure,this paper proposes a structural model based on Motif-subphrase-phrase.The model uses the seq2 seq model and GAN model to simulate the steps composers compose music:Building subphrase from motif,phrase from subphrase and chords orchestration from phrase so that the automation-generated song can be structured in a complete way,and features a motif and a distinct theme.3.Based on the melody created on step 2,chords are automatically orchestrated to add more color to the song.In the current automatic orchestration approaches,the chords are monotonous,and even the chords banned in music theory are used.In order to solve this problem,this paper employs the seq2 seq model and introduces the attention mechanism to learn the orchestration style of Johann Sebastian Bach,making the orchestration sound more natural and more colorful.Compared with other existing methods,the songs created by the automatic composition model in this paper have the advantages of high chord accuracy,clear song structure and less machine creation features.The composition model and steps provided are an effective way of automatic composition.