Application of Deep Learning Technology in Music Score Recognition and Composition

Music enriches people's lives, whether it is the soundtrack in TV dramas or movies or the music they enjoy every day, it is inseparable from people's daily life. In recent years, with the development of DL (Deep learning) method to a new height, speech recognition, speech synthesis, natural language understanding and related technologies have also made positive research progress. In this paper, the application of DL technology in music score recognition and composition is studied. Based on the framework of DL, this paper proposes a multi-scale residual CNN (Convolutional Neural Network) and BiLSTM (Bi-directional Long Short-Term Memory) to identify music scores. In order to simulate the non-linear composition mode of musicians, this paper also uses the pseudo Gibbs sampling method to generate music, and finally uses GAN (Generative Adversarial Networks) to determine whether the data distribution of the generated music and the real music is similar, and finally generates music. The research results show that the proposed method achieves good results in both symbol error rate and sequence error rate. The music generated by this method is higher than the music generated by LSTM and FNN(feedforward neural network) after regular training in five aspects, although it is slightly lower than the real music, but it is also close.