Deep Learning Models for Melody Perception: An Investigation on Symbolic Music Data

We investigate the deep learning approaches on the melody extraction problem on symbolic music data. Specifically, we compare two different approaches: the first one employs recurrent neural networks (RNN) by considering melody extraction as a sequence prediction problem, while the second employs fully convolutional networks (FCN) by considering it as a image semantic segmentation problem. Both methods are tested against a MIDI dataset with melody tracks acting as ground truth. A more challenging case that the melodies are shifted by one octave is also considered. Evaluation results show the advantage of the semantic segmentation approach in terms of the accuracy.