Short-Time ICA for Blind Separation of Noisy Speech

Noisy ICA model offers a viable solution for blind speech signal separation under real life scenarios, but it also creates new problem which need to be tackled, namely: estimating of the de-mixing matrix with non-invertible model. In this paper, a new algorithm based on short-time ICA is proposed for accurate estimation of the de-mixing matrix in the presence of noise. Without any assumption or prior knowledge of the noise covariance, the derivation of the proposed algorithm solely depends on the fact that the distribution shape remains pretty much unchanged although the estimated optima of the de-mixing matrix drift with noise contamination. In this work, we have set up a criterion on the distribution decision and signal-frame-selection for more accurate estimation. Experiment on Mandarin noisy speech showed quite satisfactory performance of matrix-estimation, thereby allowing better blind speech signal separation under noisy environment.