Segmentation of Folk Songs with a Probabilistic Model

Structure is an important aspect of music. Musical structure can be recognized in different musical modalities such as rhythm, melody, harmony or lyrics and plays a crucial role in our appreciation of music. In recent years many researchers have addressed the problem of music segmentation, mainly for popular and classical music. Some of the more recent approaches are Mauch et al. (2009), Foote (2000), Serr`a et al. (2012) and McFee & Ellis (2014). Last three are included in the music structure analysis framework MSAF Nieto & Bello (2015). None of the mentioned approaches however, addresses the specifics of folk music. While commercial music is performed by professional performers and recorded with professional equipment in suitable recording conditions, this is usually not true for folk music field recordings, which are recorded in everyday environments and contain music performed by amateur performers. Thus, recordings may contain high levels of background noise, equipment induced noise (e.g. hum) and reverb, as well as performer mistakes such as inaccurate pitches, false starts, forgotten melody/lyrics or pitch drift throughout the performance. One of the most recent approaches which addressed folk music specifics was presented by M¨uller et al. (2013). The approach was designed for solo singing and was evaluated on a collection of Dutch folk music by Muller et al. (2010). In our paper, we present a novel folk music segmentation method, which also addresses folk music specifics and is designed to work well with a variety of ensemble types (solo, choir, instrumental and mixtures).