Accompaniment Generation Based on Deep Learning and Genetic Algorithm

To generate accompaniments when given vocals, deep learning and genetic algorithm models were used to do prediction. For deep learning, a LSTM model was used. Softmax was used as an activation function and loss was determined using categorical cross entropy. The model trained with 500 epochs and 0.01 learning rate got the best accuracy 82.6. For genetic algorithm, functions were defined based on chord designs. The model was built by six steps: encoding, create initial population, evaluate individual fitness, selection, crossover, and mutation. The model with 5000 generations and 1024 population size got the best result.