Research And Application Of Biaxial LSTM Neural Network And Chaos Theory In Music Generation System

With the rapid development of modern science and technology in recent years,increasingly importance has been attached to the technology and applications of artificial intelligence.As an important research field of artificial intelligence,algorithmic composition or automatic music composition system has already obtained a lof of research achievements.Among the methods of music modeling,artificial neural network has been widely used for its strong flexibility and scalability.Aiming at the situation that the existing artificial neural network model could not maintain the invariance of the training data set,a biaxial LSTM neural network structure is proposed in this paper.Meanwhile the chaos theory is introduced as the inspiration part of the music creation process,and a hybrid system of music generation is designed.The feasibility and practicability of the system are illustrated by the experiments.The main work of this paper is as follows:Firstly,a design of a biaxial LSTM neural network for training a multi-part music data set is presented.The function of this model is generating multi-part music phrases,it can also provide an effective chord progression for the mono mode.Inspired by the structure of convolutional neural networks which is used to keep multiple invariant structures in the image processing,the biaxial LSTM neural network was designed to keep the invariant properties of multi-part training data both in time and music transposition.The accuracy of this neural network model in the prediction of the note sequence is compared by experiment.The advantages of biaxial LSTM neural network while it works in the situation of the training data was not transposed into the same tonality are discussed.At the same time,we use the model to generate a rich musical structure features of the multi-part music phrases.Secondly,a melody generation algorithm based on chaos theory is proposed.The algorithm maps the solutions of the chaotic dynamical system to the pitch and duration of the notes in the predefined scale,and generates the melody with chaotic properties through normalization and quantization.Then,the results of the continuous system and the discrete system model are compared respectively.The results are evaluated from the subjective and objective aspects.The experimental data are analyzed to obtain the influence of the dynamic parameters of the chaotic attractor on the generated melody.In addition,the result is treated as one of the sources of the training data set of the LSTM network.Thirdly,we designed a controllable music generation system which integrates the chaotic module and the LSTM module.LSTM module has mono and poly mode.The polyphonic mode is used to train and generate polyphonic polyphonic music and chord data,and the result of the latter will be provided to the mono mode as one of the conditions to keep the note invariant.In the mono mode,the LSTM module accepts the melody set from the chaotic module with chaotic characteristics,and also accepts the user input tone melody training set.Then,the system calculates the Euler degree of melodiousness of the result,and feeds it back to the input terminal,compares it with the reference melody value set by the user,then amplifies the deviation value and enters it into the chaotic module,so that it can adjusts the chaotic system parameters.In addition,by adjusting the ratio of the two training sets,we can also change the characteristics of the melody results.Finally,the performance of the system is evaluated by the experimental results,and the validity of the proposed model and the algorithm is illustrated.