Phrase-Level Audio Segmentation of Jazz Improvisations Informed by Symbolic Data

Computational music structure analysis encompasses any model attempting to organize music into qualitatively salient structural units, which can include anything in the heirarchy of large scale form, down to individual phrases and notes. While much existing audio-based segmentation work attempts to capture repetition and homogeneity cues useful at the form and thematic level, the time scales involved in phrase-level segmenation and the avoidance of repetition in improvised music necessitate alternate approaches in approaching jazz structure analysis. Recently, the Weimar Jazz Database has provided transcriptions of solos by a variety of eminent jazz performers. Utilizing a subset of these transcriptions aligned to their associated audio sources, we propose a model based on supervised training of a Hidden Markov Model with ground-truth state sequences designed to encode melodic contours appearing frequently in jazz improvisations. Results indicate that representing likely melodic contours in this way allows a lowlevel audio feature set containing primarily timbral and harmonic information to more accurately predict phrase boundaries.