The invention discloses a low SNR (signal to noise ratio) speech noise reduction method. The method comprises the following steps of: framing an input voice signal, multiplying each frame of the input voice signal into a Hamming window, carrying out time-frequency transformation and calculating spectral power distribution of the signal; according to a state of the received signal, judging to carry out gain oscillation detection on the received signal and updating bottom noise spectral power distribution; after calculating spectral distribution, inspecting an SNR and carrying out MMSE (Minimum Mean Square Error) noise suppression; calculating a frame SNR and storing and updating the frame SNR in a certain recent time; carrying out spectrum enveloped multistate transfer, and according to a state output of the state transfer, judging whether the input signal is speech or noise; carrying out frequency-time transformation and window overlapping on the signal subjected to noise reduction, carrying out speech head protection on the output signal, and according to a silence detection result, outputting the speech or silenced voice subjected to noise reduction. According to the invention, the short-wave communication noise can be effectively inhibited, quality of communication voice is improved, and effectiveness of speech noise reduction is technically ensured.