Pitch Contours Curve Frequency Domain Fitting With Vocabulary Matching Based Music Generation

In recent years,with the rapid development of computer hardware technology and the wide application of deep learning algorithms,the neural network-based method has become the research hotspot of machine learning.Through deep neural network,music generation,style transformation and music arrangement and other music operations can be realized.However,with the increasing complexity of the task,the following problems have emerged in the current methods of music generation: 1.In the long-term music generation,the style pattern often jumps and cannot complete the emotional narrative structure;2.The long-term structure of the generated music is not ideal,with a certain amount of repetitive fragments and meaningless chord;3.Lack of creativity in the generated music,resulting in irregular mixing of different melodies.Therefore,on the basis of the existing work,this paper studies the music generation and proposes a music generation method based on pitch contour curve.Based on the music generation method of pitch contour curve,this paper focuses on the long-term structure and carries out the following work:This paper presents a new method to describe the long-term structure of music melody.Based on the frequency domain decomposition of melody pitch contour curve,this method studies the frequency domain characteristics of melody pitch contour curve from the intensity distribution of frequency domain sequence to the influence of frequency domain sequence on melody pitch contour curve,and constructs a data format to describe the long-term structure of music melody through the studied characteristics.By keeping the low frequency part of the frequency domain sequence of the pitch contour curve,this special data structure format can accurately describe the long-term structure of the music melody in a certain length,which is convenient for the computer to process the long-term structure of a melody.In this paper,a long-term structural fitting neural network based on pitch contour curve is constructed.Based on the proposed method of describing long-term structure,the network takes the cyclic neural network as the main frame to generate the desired long-term structure of music.At the same time,aiming at the problem of tonality of generated music,this paper establishes a method of vocabulary matching and selection based on directed graph,and realizes the matching and replacement of melody vocabulary according to context relations and tonality restrictions.The above neural network can realize the fitting of the pitch contour curve,and can perform the fitting operation under the limited conditions based on the given label.At the same time,the word matching selection method based on the directed graph can realize the word matching operation based on the matched contour,context and long-term structure conditions,further optimize the detailed performance of the pitch contour curve,and participate in the final music generation process.Based on the long-term structure fitting neural network and word matching selection method constructed above,this paper designs a music generation algorithm based on the frequency regionality of pitch contour curve.Experiments show that the algorithm can improve the long-term performance of generated music and generate corresponding contour curves and melodies under the given long-term structure description.