Incremental bayesian audio-to-score alignment with flexible harmonic structure models

Music information retrieval, especially the audio-to-score alignment problem, often involves a matching problem between the audio and symbolic representations. We must cope with uncertainty in the audio signal generated from the score in a symbolic representation such as the variation in the timbre or temporal fluctuations. Existing audio-to-score alignment methods are sometimes vulnerable to the uncertainty in which multiple notes are simultaneously played with a variety of timbres because these methods rely on static observation models. For example, a chroma vector or a fixed harmonic structure template is used under the assumption that musical notes in a chord are all in the same volume and timbre. This paper presents a particle filterbased audio-to-score alignment method with a flexible observation model based on latent harmonic allocation. Our method adapts to the harmonic structure for the audio-toscore matching based on the observation of the audio signal through Bayesian inference. Experimental results with 20 polyphonic songs reveal that our method is effective when more number of instruments are involved in the ensemble.