With the proliferation of portable devices, speech recognition, especially name, address and command recognition on these devices is a topic of growing relevance. A mandarin short phrase recognition system is introduced in consideration of the limited resources and calculation ability of portable devices. A fixedpoint front-end is developed, discrete hidden Markov model is employed for acoustic modeling, and a SNR based likelihood weighting method is proposed to improve the noise robustness of the system. The memory size of the model set is 269kB, the decoding time is 0.89 times of the speech duration, and the method for robustness gives a relative 15.2% word error rate reduction in a complex practical environment with both channel distortion and non-stationary noise presence.