Abstract: In order to improve the robustness of fractal endpoint detection algorithm and make it suitable for more types of noise, This paper propose a fractal endpoint detection algorithm based on short-term frequency domain. Frequency reflects the energy distribution of signal and speech harmonic component have strong periodicity and regularity. Therefore it can extract fractal dimension from the energy distribution in the short time frequency domain to distinguish speech and noise. Experimental results show that the proposed method converted to the short-term frequency domain to extract fractal dimension has better robustness. It can effectively detect noise which has strong periodicity and regularity in the time domain, such as tanks noise.