This paper deals with speech enhancement achieved by thresholding the Discrete Cosine Transform coefficients of noisy speech. The sequential Monte-Carlo algorithm is used to approximate the a posteriori threshold distribution, thus the minimum mean square error estimate of the thresholded speech samples over each frequency bin are deduced. The a priori distribution of the time-frequency varying threshold is adopted as the importance density and the Gaussian random walk is used to model the threshold particle mutation. Experiences with additive white Gaussian and car noises shown the improved performance of the proposed method, compared to current speech enhancement algorithms