Abstract:
"Considering music is a complex language, classical music generation can be a challenging task which makes it harder to generate human-sounding music using artificial intelligence. Chord progression prediction can help producers find their inspiration to create melodies by giving suggestions for chord progressions.
In mitigating the problem, The Chorder has employed an approach based on deep learning, where multi-output hybrid Convolutional Neural Network (CNN) and Long-Term Short Memory (LSTM) architecture were used to train and generate music, with inputs as notes on each MIDI excerpt and outputs being the main characteristics of a musical note. The generated music is then used to predict chord progressions by extracting chord information.
The Chorder can generate music from an input, which could be considered continuing the existing note relationships to an aesthetically pleasing sound with less computational cost than the existing models. As most music generation models use loss functions for evaluations, The Chorder was evaluated with MSE loss functions and garnered a value of 0.75, which was further improved with hyperparameter tuning techniques. While the model can infer many chord progressions, the number of notes per time step remains high, which can be considered a limitation of the current system.
"
