Jazz Music Generation Based on Grammar and LSTM

There are many features that make music composition a challenging task for computer science. many people are trying to generate music in different methods which generated music do not match music rules. This paper uses music theory grammar combined with LSTM neural network to generate jazz music. We use the interval, duration, and note category information as the input data of LSTM model what parsed from the midi files. The neural network generates the sequence of notes according to the transition probability, then parse it according to the music grammar. Through improve the music grammar in the aspect of parse note sequence our system can generate music that matches music rules. we compared the effect of generated music with some popular models. Experiment results shows that the LSTM model generated music is better, Improved music grammar can generate better music than original music grammar.