Background noise is one of the noise sources of communication system, speech enhancement can reduce or eliminate noise interference, and improve speech intelligibility. Speech power spectrum is adopted orthogonal multi window spectrum estimation for smoothing processing, to reduce music noise from spectrum subtraction under the complicated noise environment. The loss of information and estimation have effectively reduced. Using adaptive spectral subtraction coefficient to adjust spectrum gain and floor to control the residual noise, using the optimized IMCRA algorithm for noise update to speech and silence, the enhanced speech signal has acquired by spectral subtraction and waveform reconstruction after the deep neural network trained data. The simulation results show that the noise reduction effect of speech is good, speech intelligibility is well.