Research And Implementation Of Background Music Generation System For Short Video Based On Neural Network

With the popularity of short video platforms,more and more Internet users create short videos for sharing.As an indispensable part of short videos,background music plays a very important role in accentuating the atmosphere of videos and expressing users’ emotions.However,background music of short video platforms often involves copyright issues.Many background music is repeatedly used without new ideas.And the existing music generation algorithms mostly focus on generating music directly,which is difficult to deal with different emotions of the music according to user’s demands.When generating multi-track music,there are also some problems such as the generated music has a short duration,the algorithm is difficult to deal with the long-term structure of music.To address these shortages,this thesis designs and implements an easy-to-use platform for generating multi-track background music.Users can generate and download the music freely with the desired style and length by this platform.This thesis divides the work of generating music into two steps.Firstly,on the issue of generating single-track main melody,this thesis analyzes the advantages and disadvantages of the original LSTM network and Lookback mechanism.Based on this,in order to import the long-term structural information of music and the emotional characteristics of different music styles,a model called LB-Attention is proposed.Note position information and an attention mechanism are added to the model.Secondly,on the issue of generating corresponding accompaniment music based on the main melody,this thesis analyzes the existing methods of generating multi-track music,and proposes an accompaniment generation model with the architecture of encoder-decoder,which imports the characteristics of the melody sequence and the inter-track information between accompaniment tracks.Finally,based on the background music generation model,an online platform for generating background music is implemented,which provides a concise and clear user interaction interface.This thesis illustrates the ideas and the improvement of the model by theoretical analysis.According to the comparative experiments with existing algorithm and models on the data index and manual evaluation,this model can achieve better results in generating background music for short videos.With the generating platform,users can easily set the parameters of music generation,as well as listen to and download the generated background music.