In order to achieve the improved effectiveness of speech enhancement under low-SNR circumstance and the robustness of the algorithm, this paper puts forward a new speech enhancement algorithm which is based on wiener filtering algorithm combined with speech endpoint detection algorithm on account of frequency domain features. The endpoint detection algorithm adopts the ratio between spectrum entropy for wavelet packet ERB sub-band and energy entropy for improvement of frequency domain. Therein, the spectral entropy of wavelet packet ERB sub-band takes masking properties of human auditory and the difference between speech and noise signal frequency distribution into account; the frequency-domain energy takes advantages of the energy difference between voice-frames and non- voice-frames. In addition, the wiener filtering algorithm acquits real time data and uses the new parameters to distinguish voice segment and no-voice segment where noise spectrum is updated smoothly. At last, the experimental results demonstrate that the endpoint detection algorithm can be able to effectively distinguish between speech segments and no speech segments, leading to the improvement of speech enhancement in the case of low SNR and the guarantee of robustness as well as real-time of the algorithm. In contrast with the other two algorithms, the new approach to speech enhancement has a better effect.