Research And Prototype Implementation Of Music Style Recognition And Generation Technology Based On Deep Learning

Using computers to automatically generate different styles of music is a hot topic in music information retrieval and music production.More and more researchers are investing in the field of automatic music generation.At present,music generation has been used by many composers.In recent years,due to the fast development of deep learning and machine learning,as well as the high-speed improvement of computer hardware and software,it has laid a good foundation for the automatic generation of different genres music.Before that,most researchers used the deep learning network to classify and recognize music genres.Now,more and more researchers begin to use deep learning for music generation,so it is very meaningful to investigate different genres of music generation.Because LSTM network has a good effect in dealing with long time series problems,this thesis’ s author designs and implements an algorithm model which can generate many kinds of genres of music by using the relevant deep learning knowledge on the premise of having a certain understanding of LSTM network.The main research contents of this thesis include the following aspects:1.The pre-processing of music data includes the separation of tracks and the acquisition of music features which are timbre,tone and velocity,by using track stitching techniques.At the same time,the music data is processed quantitatively,and the formats of input data and output data are designed.2.Based on LSTM network,the music genre style recognition and generation network are designed separately.All sub-networks share the interpretation layer,which can greatly reduce the learning of model parameters and improve the learning efficiency.Each sub-network analyses different style of music,while the whole network achieves the function of multitask processing at the same time.3.In the parameter selection of the network,the influence of the number of hidden layers and the number of neurons in each layer is compared by experimental method.Finally,the optimal network parameters are found.At the same time,how to select other parameters is also explained.4.The network parameters are updated by forward and backward propagation methods.In order to optimize the network,dropout coefficient is added and the optimaldropout coefficient is determined on experiments.The matrix containing music features is generated using test data,then it is converted into music that can be played by scripting.During the experiment,the genre data from GANT is used to test the model of style recognition and generation.By analyzing the spectrum and sound spectrum of the generated music sequence and the original music sequence,it shows that the network has good performance in different genres of music generation.At the same time,the advantages of this method are illustrated by comparing the effect of using RBM method and the method proposed in this dissertation.