Submodular feature selection for high-dimensional acoustic score spaces

We apply methods for selecting subsets of dimensions from high-dimensional score spaces, and subsets of data for training, using submodular function optimization. Sub-modular functions provide theoretical performance guaran-tees while simultaneously retaining extremely fast and scal-able optimization via an accelerated greedy algorithm. We evaluate this approach on two applications: data subset se-lection for phone recognizer training, and semi-supervised learning for phone segment classification. Interestingly, the first application uses submodularity twice: first for score space sub-selection and then for data subset selec-tion. Our approach is computationally efficient but still consistently outperforms a number of baseline methods. Index Terms — feature selection, Fisher kernel, acous-tic similarity, graph-based learning, submodularit