Abstract

Speech enhancement is essential for obtaining robust communication between man and machine, and machine and machine, in noisy environments. Filtering techniques based on Wiener's filtering, which signi?cantly goes back to the theory of optimal linear estimation, are still widely popular because of their theoretical optimality for minimizing mean square error (MSE) and their computational economy. In this paper, a thorough treatment of Wiener filtering is given, including treatment of its theoretical derivation, of actual short-time Fourier transform (STFT) realizations, of methods for estimating noise power spectral density (PSD), of a method for decision-directed a priori SNR estimation, as well as methods for reducing perceived musical noise. An algorithmic description is given of the procedures employed along with certain suggested parameters. A set of representative experimental results (using standard tests and corpora and types of noise) is given in order to illustrate typical performance in low to-moderate SNR. It is shown vs. spectral subtraction and Kalman filtering that Wiener based enhancement presents a favorable trade-off between improvement of speech intelligibility (as measured by STOI), improvement of perceptual quality (as measured by PESQ), and under all cost of calculation. Finally, practical aspects are discussed with respect to problems of real time implementation as well as directions for hybrid systems based onWiener filters combined with modern deep learning estimating techniques.