Fast model based optimization of tone onset detection by instance sampling

Abstract

There exist several algorithms for tone onset detection, but finding the best one is a challenging task, as there are many categorical and numerical parameters to optimize. The aim of this task is to detect as many true onsets as possible while avoiding false detections. In recent years, model-based optimization (MBO) has been introduced for solving similar problems. The main idea of MBO is modeling the relationship between parameter settings and the response by a socalled surrogate model. After evaluating the points of an initial design—each point represents here one possible algorithmconfiguration—themain idea is a loop of two steps: firstly, updating a surrogate model, and secondly, proposing a new promising point for evaluation. While originally this technique has been developed mainly for numerical parameters, here, it needs to be adapted for optimizing categorical parameters as well. Unfortunately, optimization steps are very time-consuming, since the evaluation of each new point has to be performed on a large data set of music instances for getting realistic results. Nevertheless, many bad configurations could be rejected much faster, since their expected performance might appear to be very low after evaluating them on just a small partition of instances. Hence, the basic idea is to evaluate each proposed point on a small sample and only evaluate on the whole data set if the results seem to be promising.