情報処理学会 第75回全国大会講演要旨

The goal of this work is a system capable of following human accompaniment. The fundamental difficulty of score following for human accompaniment is that the score －a lead-sheet, common in popular music－ only provides chord-names. Thus, actual accompaniments have uncertain octave and timber; fluctuating rhythm and tempo. This can degrade performance in conventional score following systems. Our solution is using chroma vectors as features for a particle filter. We measure robustness to timber and tempo changes by testing our system on 9 sets of 20 songs, covering all combinations of 3 levels of accompaniment complexity, and 3 types of instruments.