A Study Of Piano Composition Based On Transformer

Deep learning and neural networks are constantly evolving and deep learning techniques are showing their potential in many areas.The use of deep learning techniques for various content generation has become a trend,with algorithmic composition being one of the popular areas of research.Algorithmic composition not only allows the general public to experience the joy of creating music,but also provides professionals with creative inspiration and greatly reduces the time required to create music.The aim of this paper is to generate music with harmony and melody using a neural network approach with strong musical feature learning capabilities.The research in this paper is based on aspects of deep learning and neural network composition.The research covers the background of algorithmic composition,basic music theory,MIDI file processing,the construction of neural networks and music generation and evaluation.The main work and innovations are as follows:(1)This paper proposes a Random-Masking Transformer(RM-Transformer)model for composing based on the Transformer network model.In the paper,we preprocess the music data and design a random mask matrix so that the model training can consider the information of future notes and improve the overall harmony and melody of the music.After comparing the experimental results,the RM-Transformer network has improved the similarity of music composed by 4% and reduced the ratio of empty bar by 3% compared to LSTM,Transformer and other network models.(2)In this paper,we have done a more in-depth study of deep learning networks.In order to further improve the harmony of the main melody and chords in the generated music,a step-by-step Grouping-Combining Transformer(GCT)network is proposed for piano composition.The whole model is divided into two modules: melody generation and chord generation.After the note melody is obtained in the melody module,the chord and melody dependencies are then learned.Through the evaluation method proposed in this paper,the music is evaluated in terms of musical pitch,chord and pitch distance.The model not only generates long sequences of music with rich musical features,but also has a great improvement in chord generation,and the GCT model has a 9% lower ratio of empty bar than BLSTM in the generation of music of 60 seconds length,and indicators such as chord-melody pitch distance are closer to the music database(3)A smart music platform is built,which is a human-computer interaction interface built using the Py Qt5 framework and My SQL database,and the interface as a whole conforms to the trend of UI interface design.The music platform is divided into functional modules such as music generation,music recognition,music retrieval,music evaluation and music visualisation.The music generation module is coupled with the composition algorithm proposed in the paper,and users can compose music and play it on the intelligent music platform,and users can also score the composed music through the music evaluation function.The functional modules on the Smart Music Platform can initially meet the needs of users.