Speech enhancement using a-priori information

In this paper, we present a speech enhancement technique that uses a-priori information about both speech and noise. The apriori information consists of speech and noise spectral shapes stored in trained codebooks. The excitation variances of speech and noise are determined through the optimization of a criterion that finds the best fit between the noisy observation and the model represented by the two codebooks. The optimal spectral shapes and variances are used in a Wiener filter to obtain an estimate of clean speech. The method uses both a-priori and estimated noise information to perform well in stationary as well as non-stationary noise environments. The high computational complexity resulting from a full search of joint speech and noise codebooks is avoided through an iterative optimization procedure. Experiments indicate that the method significantly outperforms conventional enhancement techniques, especially for non-stationary noise.