Code for reproducing the playing technique recognition system (Section 5.5.2) in the thesis:

C. Wang. "Scattering Transform for Playing Technique Recognition", PhD thesis, Queen Mary University of London, 2021.

This work proposes two scattering transform variants, the adaptive scattering and the direction-invariant joint time--frequency scattering (dJTFS). The code for extracting these features is build upon the ScatNet toolbox. We organise the implementation by four stages:

Download the complete CBFdataset directly from zenodo.org/record/5744336.

Detect the fundamental frequency (F0) as the decomposition trajectory.

We extract the AdaTS+AdaTRS feature and the dJTFS-avg feature using by calling Matlab as a Python subprocess. The AdaTS+AdaTRS is the concatenation of adaptive time scattering (AdaTS) and the adaptive time--rate scattering (AdaTRS) while the dJTFS-avg is dJTFS obtained by applying average pooling to the direction variable of the joint time--frequency scattering.

With the scattering features extracted, we use a support vector machine classifier to label the playing techniques.

Any questions/bugs, please feel free to contact the author at [email protected].

The thesis template is built upon William J. Wilkinson's thesis.