2009 Research Summary

Ensemble Feature Selection for Multi-Stream Automatic Speech Recognition

View Current Project Information

David Gelbart and Nelson Morgan

German Federal Ministry of Education and Research

Automatic speech recognition systems often use an ensemble of classifiers approach, in which a set of classifiers that each use a different feature extraction method are run in parallel and have their decisions combined. However, past work on how best to assign features to classifiers in such systems has always dealt with groups of features rather than individual features. We have demonstrated the feasibility and effectiveness of assigning individual features using an automated search procedure. We have also investigated how the effectiveness of automated search depends on the amount of data used to guide the search, and we are now investigating how the size of the ensemble interacts with the performance of the search.