The speech recognition system, as one of the steering control components of the Manual Wheelchair Automator (MWA), is designed to benefit the end users who have lost control of their upper extremities. An alternative joystick steering method is incorporated into the control system to provide the users with more options. In particular, the speech recognition system consists of two parts: the first part is a small vocabulary training section that constructs a model, and the second is a speech recognition section that uses this model. This speaker independent, discrete word speech recognition system is implemented on an 80 MHz, 32 bit micro-controller. Much research has been done on the existing techniques on implementing speech recognition. The result is that a Hidden Markov Models (HMMs)-based method with Viterbi re-estimation and Forward- Backward Model are both selected for model construction based on its reliability and relative efficiency. Our experimental results indicate that Forward-Backward re- estimation provides a better performance. Also, the speech recognition section adopts the Viterbi Algorithm to measure the likelihood of speech input to the model of the command words. The algorithm, design of the system, experimental data, and future improvements are presented in detail.