Multi-emotional Music Generation Algorithm Based On Double Encoding Of Chord And Melody

Music is an art,and with the development of the deep learning,it has become a research hotspot to enable computers to achieve automatic music composition.The complete composition is usually divided into three stages,corresponding to three tasks in music generation,which are the symbolic domain music generation task,the expressive music generation task by adding advanced features such as emotion and style,and the audio synthesis task.Currently,deep learning has been widely used in these three music generation tasks,but there are still some problems in each task that have not been solved yet.In the symbolic domain generation task,most of the studies on chords are limited to common triads,which are simple and lack of diversity,and most of the studies are more focused on melody generation and not deep enough on chord generation.In the expressive music generation task,music generation with multi-emotion has been the focus of research,and most of the current multi-emotional music generation tasks use datasets with emotional labels to train the models,but it is timeconsuming and expensive to add emotional labels to the data,and the small datasets make the models cannot train well.In addition,emotion is subjective,and there may be large differences in perceptions between different markers and raters,all of which lead to poor training and learning of the model on a manually labeled music emotion dataset.In the audio synthesis task,the quality of the synthesized audio is always a problem to be solved.The main research work of this paper addresses the problems in the three stages of music generation as follows:(1)Music generation model based on double coding of chord and melodyAiming at the problem in symbolic domain,a music generation model based on chord and melody double coding is proposed in this paper.Firstly,a new chord representation consisting of four One-hot vectors is proposed to represent more kinds of chords such as seventh chord.Then,this paper makes the model learn the musical characteristics and structural relations to generate more musical chords by double-coding the chords and melodies.In addition,two Selfattention layers are added to enhance the structure of the model and improve its performance.(2)Multi-emotion music generation model with controllable rhythm and modeTo address the dataset problem in expressive music generation tasks,a multi-emotional music generation model is proposed,it can control the rhythm and tonal,which is trained without relying on music datasets with emotional labels,which solves the problem in multiemotional music generation tasks.Firstly,this paper uses pitch histogram vectors to control the tonal of the music,and realizes the control of C major and A minor.Secondly,a new chord rhythm representation and a beat representation are proposed in this paper,and the chord representation in the first stage are combined with them to control the density of chords,i.e.,the rhythm.Finally,the music can be generated with different emotions by adjusting the rhythm and tonal.(3)Audio synthesis model based on Performance NetThe Performance Net model not only achieves the mapping from symbolic domain to audio domain,but also learns to synthesize audio in an end-to-end manner.Based on this,the loss function of this model is improved to achieve the improvement of the model performance and thus enhance the quality of the final synthesized audio.Since the influence of parameters on the quality of synthesized audio cannot be ignored in the audio combining task,this paper improves the quality of synthesized audio by selecting the optimal parameters through experimental comparison.