Optimized Active Learning Strategy for Audiovisual Speaker Recognition

摘要

The purpose of this work is to investigate the improved recognition accuracy caused from exploiting optimization stages for tuning parameters of an Active Learning (AL) classifier. Since plenty of data could be available during Speaker Recognition (SR) tasks, the AL concept, which incorporates human entities inside its learning kernel for exploring hidden insights into unlabeled data, seems extremely suitable, without demanding much expertise on behalf of the human factor. Six datasets containing 8 and 16 speakers' utterances under different recording setups, are described by audiovisual features and evaluated through the time-efficient Uncertainty Sampling query strategy (UncS). Both Support Vector Machines (SVMs) and Random Forest (RF) algorithms were selected to be tuned over a small subset of the initial training data and then applied iteratively for mining the most suitable instances from a corresponding pool of unlabeled instances. Useful conclusions are drawn concerning the values of the selected parameters, allowing future optimization attempts to get employed into more restricted regions, while remarkable improvements rates were obtained using an ideal annotator.