Robust speech enhancement using multi H∞ filters

Since the speech enhancement algorithm based on a Kalman/Wiener filter requires a priori knowledge of the noise statistics and is focused on the minimization of the variance of the estimation error of the speech signal, a small estimation error on the noise may lead to a large estimation error. However, an H_∞ filter does not require any assumptions and a priori knowledge on the noise statistics and it searches the best estimated signal that has the minimum error from all the estimated signals applying least upper bound, so it is more robust to a change of noise than the Kalman/Wiener filter. We propose a speech enhancement method based on multi H_∞ filters. First, HMM parameters are estimated with the training data. Second, the speech signal is filtered with a fixed number of H_∞ filters. Finally, the estimated clean speech is obtained from the sum of the weighted filtered outputs. Experimental results shows about 1-2 dB SNR improvement with increment of computation compared with the conventional Kalman filtering method.