IJMSTA - Vol. 1 - Issue 1 - January 2019
A Greedy Approach to Music Voice Separation Using Hidden Markov Models
Authors: Edson E. da Silveira, Shigeki Sagayama
Abstract - This research, in the area of music information retrieval, describes a probabilistic method of separating notes into different voices for music pieces encoded as MIDI files. The proposed solution represents the music as a sequence of chords that are analyzed individually in a linear fashion. The problem of voice assignment is modelled as an HMM problem at each chord. Given a list of all previously given assignments, the model suggests the most likely voice assignments for the current chord. The model was tested on a dataset of 111 J. S. Bach pieces with F-measure scores for J. S. Bach's 15 Inventions (99.17) and J. S. Bach's The Well-Tempered Clavier 48 Preludes (95.28) and 48 Fugues (96.97) being reported. The results show improvement over currently available methods on the same dataset.
Keywords: Dynamic Programming, HMM, Voice Separation, Music Information Retrieval