Adaptive Learning With Surrogate Assisted Training Models for Acoustic Source Classification

This article presents an adaptive learning solution for the selection of surrogate assisted training models. The main issue is to improve classification of acoustic sources considering that few labeled data are available. In this proposal, acoustic models are initially obtained from real signals. Surrogate models are then applied to assist the original training procedure and achieve improved classification accuracy. Learned models are defined according to the discrimination power among audio classes. Results show that the learning procedure leads to substantial accuracy gain in classification experiments. The proposed solution is also evaluated as a pre-learning step for a dictionary learning algorithm. In this scenario, the average accuracy is improved for a highly nonstationary and a stationary acoustic source.