Font Size: a A A

Automatic Composition Of Guzheng Music Based On Long Short-term Memory And Deep Q-learning Network

Posted on:2023-05-29Degree:MasterType:Thesis
Country:ChinaCandidate:S L ChenFull Text:PDF
GTID:2545306830954459Subject:(degree of mechanical engineering)
Abstract/Summary:PDF Full Text Request
With the rapid development of artificial intelligence technology in the field of modern science,foreign researchers have begun to apply the deep learning methods in artificial intelligence technology to music creation.In domestic music creation,especially the automatic composition of Chinese national music,due to the small amount of national music data,the difficulty of audio format conversion,and the difficulty of restoring the characteristics of national music,few people are still involved.As a famous brand of traditional Chinese music,national music may inject new vitality and vitality into it if artificial intelligence technology is used.In this paper,Guzheng is used as a representative of Chinese national musical instruments,the Long Short-Term Memory(LSTM)network and the Deep Q-learning Network(DQN)are used to establish an automatic composition system for Guzheng music.Firstly,this paper studies the music theory knowledge of Guzheng music,organizing and collecting the data of Guzheng music,and converts it into music format(MIDI format)that can be recognized and processed by computer.In this paper,a self-similar matrix is used to process Guzheng music and extract the chorus of music,and the extraction accuracy is as high as 95%.This result is used as one of the sources of neural network data set for training,emphasizing the importance of chorus melody.Secondly,model training is performed on the basic elements of music.The generation of the melody part of the Guzheng is using the LSTM model,which made the automatically generated melody have coherence,and the DQN model is used for the generation of Guzheng music theory and techniques,which makes the final generation effect more characteristic of Guzheng.According to the modeling results,the accuracy of the model is as high as 86%,the loss is as low as 0.03,and the average gain of the reinforcement learning model is 0.98.Then,in the post-processing link of Guzheng music,the MIDI music generated by the model is converted into Guzheng timbre,through the spectrogram showed that the conversion link has a good effect,which is closer to the real Guzheng playing effect.Finally,through the accuracy of note prediction,the detection of the proportion of notes,the comparison of Guzheng music characteristics,the effectiveness of the before and after processing links are objectively evaluated,they can be obtained that the accuracy rate of note generation is more than 70%,the proportion of notes is more uniform,the repetition fragment decreases by 40%,and the technique increases by 30%.At the same time,in order to reflect the public adaptability of automatic composition,amateur and professional music critics are invited to subjectively evaluate the generated Guzheng music and sample music,which shows that the music generated by the model in this paper scores the highest.Guzheng automatic composition system includes the pre-processing link,the LSTM and DQN model training and post-processing link.This paper completes all the work of Guzheng automatic composition for the first time,the effect of generation is good and comprehensive considers Guzheng characteristics.It can extend to the whole national music work and promote the development of national art creation in the future.
Keywords/Search Tags:Guzheng music, Automatic composition, Long Short-Term Memory Network (LSTM), Reinforcement learning network (RL)
PDF Full Text Request
Related items