Classification of time delay estimates for robust speaker localization

Abstract

This paper proposes a solution to the problem of robust speaker localization under adverse acoustic conditions. The approach is based on the classification of time delay estimates. Two classification techniques are investigated in detail: maximum likelihood (ML) classification and classification based on histogram comparison. Their performance under adverse acoustic conditions is compared to outcomes obtained with the traditional approach which uses time delay estimates directly to infer speaker positions. Experiments indicate that the ML classification method provides little improvement over the traditional method. On the other hand, using histogram classification, we can improve the probability of correct speaker localization by more than 60% compared to either the traditional approach or the ML classification technique.