Generating Creative Classical Music by Learning and Combining Existing Styles

The existing music creation system(s) only prioritise accuracy, that is, producing a sound that closely resembles actual music. Their primary concern is a criterion of legitimacy; creativity and the calibre of the music produced are not given much thought. A system should be able to produce music of a particular calibre, style, or even combinations of the two if it can intelligently model both composers and existing composition styles. As a result, this gives an AI agent some degree of combinational inventiveness. A framework is presented wherein music can be generated by encapsulating the structural styles of diverse composers through the utilisation of a sequence model comprising a Generative Adversarial Network (GAN) and Long Short-Term Memory Recurrent Neural Network (LSTM RNN). Western classical music is our main priority, but we also plan to include other genres. We use the Python module music21 to interpret the MIDI audio files in the dataset. A stream of signals or messages is extracted from the MIDI file and processed sequentially. The generator network and the discriminator network, which help with style-training the generator, are trained with these sequences. The generator network creates musical samples according to the stylistic contributions of different composers after undergoing training to understand the structural patterns of musical notations. The Audio Fingerprinting method is employed to further validate the generated samples.