H-Semantics: a hybrid approach to singing voice separation

The singing voice is the most prominent content of music tracks that can be described as songs. Separation from its music accompaniment is considered highly desirable in the field of music information retrieval, as it facilitates such applications as melody extraction, lyrics recognition, and singer identification. This paper presents investigations into effective methods for unsupervised separation of singing voice from stereophonic studio recordings. The work involves the introduction of two novel time-domain procedures for music pruning and the integration of each of them with frequency-domain voice isolation, which is based on the enhancement of a previously established procedure. The performance of the complete system based on each of the above music-pruning methods is analyzed and measured using a set of experimental investigations. The outcomes clearly illustrate that the effectiveness in singing voice separation can be considerably improved through the proposed approaches.