Deep net classification of drum strike location with non-uniformmembrane tension

Abstract

Tuning inaccuracies in snare drum membranes can cause inconsistent spectro-temporal changes (e.g., timbre differences) exacerbated by drumhead strike location. This experimental study assesses the potential for developing a deep net classifier for strike drum location, using a Non-UniformlyTuned Snare (NUTS) dataset, comprising 12,107 individual snare drum hits recorded in an anechoic chamber. Unique features of this dataset include its substantial size, additional novel features comprising physical attributes of the drum, and a performance-based classification schema. A series of deep net classifiers are trained using Bayesian hyperparameter optimization, on a subset of 8685 strikes of the NUTS dataset, whose acoustic features are extracted at nine different averaged time-windows(23–2000 ms). A comparison is then made against similar models supplemented by domain knowledge of 15 different combinations of non-uniform membrane tension, whose measures are recorded near the sixteen tuning lugs for each strike. Results found above chance model accuracies, and some performance benefits of including the tuning information as domain knowledge between equivalent trials. Shorter window sizes performed worse than longer time windows, and per-class F1-scores suggest complex spectro-temporal time variant changes caused by dis-uniform membrane tension, which could have implications for performers wishing to broaden the sonic capabilities of the instrument.