Abstract: For speech enhancement algorithm based on statistical model,different distribution models are corresponding to different gain function,due to the uncertainty of the speech signal,no distribution function can accurately model the speech and noise spectra distribution,so any kind of fixed reference models will have some errors.We presented a gain dictionary queries based speech enhancement algorithm,getting a dictionary gain through training the voice of a noise library using log-spectral distortion criterion,for which the input is the estimate value of a priori and a posteriori SNR.Finally,we used ITU-T P.826 PESQ,segmented SNR,total SNR and log-spectral distortion criterion to test the proposed algorithm,and compared this algorithm with Gaussian distribution model and Laplace distribution model.The experimental results show that the algorithm is better than the other algorithms,whether in stationary or non-stationary noisy environments,and musical noise and residual background noise can be well suppressed.

Key words: Speech enhancement,Dictionary queries,Decision-directed,IMCRA