An improved method for speech enhancement, which is based on particle filtering, is presented in this paper. The Rao–Blackwellized particle filter (RBPF) is used to estimate the model parameters and recover a clean speech signal. The proposed method (named DFT-RBPF) enhances the complex discrete Fourier transform (DFT) coefficients of a noisy speech signal. The real and imaginary parts are filtered separately, under the assumption of mutual independence, using a low-order time-varying auto-regressive (TVAR) process with a linear Gaussian model. The obtained results, in terms of the coherence speech intelligibility index (CSII), perceptual evaluation of speech quality (PESQ), segmental and overall signal-to-noise ratios (SNRseg, SNR), demonstrate the improved performance of the proposed method, when compared with the recent methods based on particle filters and the existing algorithms for speech enhancement.