Research on the noise adaptability of spectral subtraction enhancement method for speech communication

This paper addresses the problem of noise adaptability for a certain speech enhancement method, which is the most important pre-processing techniques in the speech communication system. Considering that the complicated and variety of the practical background noises and the convolution channel noise, the non-linear spectral subtraction (NSS) algorithms was adopted to enhancing the noisy speech signal. The noise data comes from the Noisex-92 database. Segmental signal-to-noise ratio (SegSNR) and Itakura-Saito distance (IS) are used as the objective measures to evaluate the performance of the NSS enhancement algorithm to the noisy speech corrupted by different types of noise at different SNR levels. Simulation results indicate the efficiency and adaptability of the proposed method. It shows that the colored noise is more suitable for the NSS algorithm than the white Gaussian noise.