A Novel Scheme of Speech Enhancement using Power Spectral Subtraction - Multi-Layer Perceptron Network

Abstract

A novel method for eliminating noise from a noised speech signal in order to improve its quality using combined power spectral subtraction and multi-layer perceptron network is presented in this paper. Firstly, the contaminated speech signal was processed by spectral subtraction to enhance the clean speech signal. Then, the signal was processed by a neural network using the spectral subtraction parameters and result of estimated speech signal in order to improve its signal quality and intelligibility. The artificial neural network used was multi-layer perceptron network consisted of three layers with six input and one output. The neural network was trained with three speech signals contaminated with two level white gaussian noises in SNR including 0 dB and 30dB. The designed speech enhancement was examined with ten noised speech signals. Based on the experiments, the improvement of signal quality SNR was up to 7 dB when the signal quality input was 0dB. Then, based on the PESQ score, the proposed method can improve up to 0.4 from its origin value. Those experiment results show that the proposed method is capable to improve both the signal quality and intelligibility better than the original power spectral subtraction.