Having developed about 50 years, the speech recognition (SR) technique has a wide range of applications in many fields, such as voice dialing system, digital remote control and industrial control. But the limitation of acoustic and language model is that the computer can only recognize some words or sentences. When the speech language changes, the system often gets wrong results. To address the problem above, the speech recognition system has been built on the basis of HTK as well as hidden markov model theory. Controlling the building process by code, the system can quickly generate a new recognition model when the training data and dictionary has changed.