Abstract: Due to the sharp decline in the recognition rate from noise interference in speech recognition systems, there are limitations in the spectrum estimation of speech signals in traditional robust speech feature extraction methods. In this paper, we discuss these limitations and propose a new speech feature extraction method with better robustness and recognition performance under different signal-to-noise ratios (SNRs). In the new method, spectrum estimation involves the combination of a multiple signal classification (MUSIC) method and a minimum-norm method (MNM). We conduct a verification test of the new method on mobile robot platforms and the results show that it has a higher recognition rate and better robustness in speech recognition.