Finding Repeating Stanzas in Folk Songs

Folk songs are typically composed of repeating parts stanzas. To find such parts in audio recordings of folk songs, segmentation methods can be used that split a recording into separate parts according to different criteria. Most audio segmentation methods were developed for popular and classical music, however these do not perform well on folk music recordings. This is mainly because folk song recordings contain a number of specific issues that are not considered by these methods, such as inaccurate singing of performers, variable tempo throughout the song and the presence of noise. In recent years several methods for segmentation of folk songs were developed. In this paper we present a novel method for segmentation of folk songs into repeating stanzas that does not rely on additional information about an individual stanza. The method consists of several steps. In the first step breathing (vocal) pauses are detected, which represent the candidate beginnings of individual stanzas. Next, a similarity measure is calculated between the first and all other candidate stanzas, which takes into account pitch changes between stanzas and tempo variations. To evaluate which candidate beginnings represent the actual boundaries between stanzas, a scoring function is defined based on the calculated similarities between stanzas. A peak picking method is used in combination with global thresholding for the final selection of stanza boundaries. The presented method was tested and evaluated on a collection of Slovenian folk songs from EthnoMuse archive.