Probabilistic Segmentation of Folk Music Recordings

Abstract

The paper presents a novel method for automatic segmentation of folk music field recordings. The method is based on a distance measure that uses dynamic time warping to cope with tempo variations and a dynamic programming approach to handle pitch drifting for finding similarities and estimating the length of repeating segment. A probabilistic framework based on HMM is used to find segment boundaries, searching for optimal match between the expected segment length, between-segment similarities, and likely locations of segment beginnings. Evaluation of several current state-of-the-art approaches for segmentation of commercial music is presented and their weaknesses when dealing with folk music are exposed, such as intolerance to pitch drift and variable tempo. The proposed method is evaluated and its performance analyzed on a collection of 206 folk songs of different ensemble types: solo, two- and three-voiced, choir, instrumental, and instrumental with singing. It outperforms current commercial music segmentation methods for noninstrumental music and is on a par with the best for instrumental recordings. The method is also comparable to a more specialized method for segmentation of solo singing folk music recordings.

Transcription of Polyphonic Vocal Music with a Repetitive Melodic Structure

“…Segmentation and pitch drift are obtained with a method presented in [25]. We define segmentation as a set of boundaries between repeated segments = {ω i }, where ω i represents the beginning time of the i-th segment.…”

“…The method achieves an F1 score of 0.76 on a collection of folk music recordings. For more details on the algorithm and its evaluation, we direct the reader to [25].…”

“…To be able to compare and align different parts of a song, we need to adjust the F0 estimates according to the estimated change of intonation. As pitch drift D is already estimated during segmentation [25], F0 values are adjusted in this step according to the estimated drift, as in:…”

“…Alignment is performed with dynamic time warping (DTW) over segment pairs represented by respective piano roll excerpts P ω i . We use correlation distance (one minus correlation coefficient) as our local distance measure, which has been shown to perform well for this task (as shown in [25,26]). To be more robust to individual incorrectly performed notes and other pitch fluctuations, such as vibrato or pitch changes during onset and offset, we smooth the piano roll representation with a Gaussian filter over the frequency axis prior to DTW calculation.…”