Application and performance analysis of LSTM networks in polyphonic popular music generation

Deep learning-based music generating techniques have slowly shown notable advancement in the area of popular music composition with the fast evolution of artificial intelligence technology. This work intends to look at how long-short-term memory (LSTM) networks are used in polyphonic pop music generation and their performance. An LSTM-based generative model is therefore created to properly catch the temporal dependencies in popular music and produce melodies and harmonies following the rules of music. Experimental findings indicate that, particularly in the coordination between several voices, the LSTM network can better preserve the harmony and consistency of the song when producing polyphonic music. At last, this study offers a perspective for future research considering the constraints of the present work; with the ongoing enhancement of dataset diversity and model optimisation, smart music composition will become more and more relevant in the domain of music composition.