Missing feature reconstruction methods for robust speaker identification

In this study, we propose a reconstruction method to restore the degraded features for robust speaker identification. The proposed method is based on a hybrid generative model which consists of deep belief network (DBN) and restricted Boltzmann machine (RBM). Specifically, the noisy speech is firstly decomposed into time-frequency (T-F) representations. Then ideal binary mask (IBM) is computed to indicate each T-F point as reliable or unreliable. We reconstruct the unreliable ones by the proposed model iteratively. Finally, reconstructed feature is utilized to conventional speaker identification system. Experiments demonstrate that the proposed method achieves significant performance improvements over previous missing feature techniques under a wide range of signal-to-noise ratios.