Recurrent Neural Network for MIDI Music Emotion Classification

MIDI(Music Instrument Digital Interface) music is the most widely used music standard format. This paper proposes a MIDI music emotion classification method based on RNN(Recurrent Neural Network). It uses the combination of two notes melody trend as music features to realize the classification of the five kinds of music emotion, with accuracy up to 75.4%. The dropout parameter is added to the RNN network to improve the accuracy rate of nearly 10%. Emotion tagging MIDI music on Youtube are used as the dataset. The learning rate, initial weight and bias, as well as the types of notes, are compared and analyzed between RNN, improved RNN and FNN(Forward Neural Network).