Intelligent human robot interaction are becoming popular in both industry and academia. However, amongst current techniques, speech recognition is a challenging topic, including real-time translation with high accuracy, amicability and the support for recognizing minor languages or sophisticated dialects. In this paper, we propose a human-friendly prototype deployed on NAO robots in a real-life scenario through daily speech commands and NAO would act accordingly. We primarily adopt HMM-GMM, the combination of HMMs (Hidden Markov Models) and GMMs (Gaussian Mixtures Models). The experimental results show that the proposed prototype achieves high accuracy and well-received by experiment subjects.