Adaptive Scattering Transforms for Playing Technique Recognition

Playing techniques contain distinctive information about musical expressivity and interpretation. Yet, current research in music signal analysis suffers from a scarcity of computational models for playing techniques, especially the context live performance. To address this problem, our paper develops general framework technique recognition. We propose adaptive scattering transform, which refers to any transform that includes stage data-driven dimensionality reduction over at least one its wavelet variables, representing techniques. Two features are presented: frequency-adaptive direction-adaptive scattering. analyse seven techniques: vibrato, tremolo, trill, flutter-tongue, acciaccatura, portamento, glissando. evaluate proposed methodology, we create new dataset containing full-length Chinese bamboo flute performances (CBFdataset) with expert annotations. Once trained on representations, support vector classifier achieves state-of-the-art results. provide explanatory visualisations coefficients each verify system three additional datasets various instrumental vocal VPset, SOL, VocalSet.