Active learning and semi-supervised learning for speech recognition: A unified framework using the global entropy reduction maxi