A Melody Generation Algorithm for Popular Music Based on Multidimensional Spectral Analysis and Its Innovation on Compositional Patterns

Abstract

In this paper, by performing Fourier transform on the original signal of the musical melody, obtaining its audio signal, and arranging the result to form a speech spectrogram, and utilizing multidimensional spectral analysis method for this to extract the musical melody. On this basis, a convolutional neural network model is added to extract the musical melody by means of the temporal harmonic map convolutional network model. After the pre-training of BERT model, the new musical melody is generated by Transformer, which defines the audio symbols and assists in the musical arrangement. The total number of segments above 1 point in the music clips generated using the model of this paper are 1349, 1047, and 687, respectively. More than 90% of the passages were audibly acceptable, indicating that the structure of the melodic flow generated in this paper has a high degree of similarity to the musical chords, and is subjectively recognized by the listener. The violin audio and the flute audio reached their maximum values in the first and second frames respectively, with the highest frequencies of 0.465 and 0.0198, respectively, and the violin timbre was richer.