Strategies for modeling reverberant speech in the feature domain