LSTM-MorA: Melody-Accompaniment Classification of MIDI Tracks

Many studies based on symbolic music signals require retaining only melody tracks or accompaniment tracks from musical instrument digital interface (MIDI) files. However, this seemingly simple setting often becomes a stumbling block in the first step because the MIDI format does not have any mandatory regulations for the track numbers of melody/accompaniment tracks. This study delves into the classification of melody and accompaniment parts within MIDI files, pioneering the use of long-short-term memory (LSTM) for this purpose. An LSTM network is trained to classify multivariate time series of varying lengths, representing the tracks within MIDI files as either melody or accompaniment. Experimental results of over 0.91 accuracy, precision, recall and F score reveal that our proposed methodology, LSTM-MorA (Melody or Accompaniment), could be one of the solutions for MIDI melody-accompaniment classification.