MIDI melody extraction based on improved neural network

Standard MIDI files consist of a number of tracks. Usually, one of them is melody track, and others accompaniment tracks. To recognize the melody track from multiple tracks is important for the music retrieval and other music related applications. Though lots of researchers had researched on this topic, more efficient and precise methods are still needed. An innovative method based on the improved neural network to distinguish melody from accompaniment is proposed to recognize the melody track from multiple tracks. A set of features from each track of the MIDI file are extracted. These features are input into an improved neural network classifier that assigns the probability of being a melodic line to each track. Then, the track with the highest probability is chosen as the melodic line of the MIDI file. Experiments show that when compared with other methods, the proposed approach is outperformed in recognition precision and is effective for solving the melody recognition problem.