Adaptively compensated multiband spectral subtraction for robust noise reduction

Abstract

An adaptively compensated multiband spectral subtraction (MBSS) is presented in this paper. In this research, the adaptive compensation in the MBSS utilizes artificial neural network. The purpose of this compensation is to improve the quality of speech signal after denoising of MBSS step. This compensation is calculated adaptively depend on the MBSS parameters, estimated noise, and difference between input and estimated speech signal. The neural network used was Multi-Layer Perceptron consisted of three hidden layers. The proposed neural network was trained by three speech signals contaminated by white gaussian noises with SNR 0dB and 30dB. For investigating the performance, the proposed method was tested by five noised speech signals with SNR 0dB to 10dB. The result of experiment is examined and evaluated by SNR and PESQ scores. Based on the examination, the proposed speech enhancement method exposed the better performance than the origin MBSS algorithm.