Variance Reduction for Optimization in Speech Recognition