Transliteration is a common translation method when named entities are introduced into another language. Direct orthographical mapping (DOM) approach is successfully applied in machine transliteration by segmenting a word according to syllables and then mapping them directly into target language without considering its pronunciation. The paper studies the performance of two-stage machine transliteration based on Conditional Random Fields. To reduce the amount of computation in model training, we propose an error-driven learning by dividing the training data into several groups and training the transliteration model step by step based on the error prediction data until the performance doesn't increase or the limitation of the computer. Experiments on data of NEWS2011 show that error-driven model training reduces computational complexity and saves the time of model training. Compared to the combining transliteration model, our transliteration system increases the accuracy of top-1 output with 0.06, reaching 0.652.