Maximum entropy-based reinforcement learning using a confidence measure in speech recognition for telephone speech