Music Generation using Deep Learning with Spectrogram Analysis

In the popular music market, the music producer takes a large amount of time and effort to produce one piece of decent music. It would be much more productive and efficient if music generation with the machine was made possible, thus our study. LSTM and GAN are implemented in our study. To analyze the generated music, we used spectrogram analysis. We generated several pieces of music using LSTM and GAN. From the loss curves, we see a promising generation of the two models. Analyzed with a spectrogram, we see the music generated by LSTM is more coherent, while music generated by GAN is more rhythmic. The results we have are relatively good, and the spectrogram analysis part is intelligible and clear. Despite our hardware and computing ability limitations, we still have promising results generated from our models. For future works, we might be able to generate undistinguishable music with little effort.