Time-domain speech enhancement using neural networks

Abstract

The scope of this project covered the objective of filtering the background noise in speech signal using a Neural Network (NN) while reducing the loss of speech content. More appropriately, it should be called as an Artificial Neural Network (ANN) as computer simulation was done to the Neural Network. This is similar to the Neural Network of the human brain. Neural Networks are made up of the biological neurons while Artificial Neural Networks are built based on artificial neurons which can be implemented on an electronic device like a computer. Both NN and ANN share the same concept of imputting information into the network so that they will able to train on their own to produce the desired results. In the Time-Domain NN, input corrupted noisy speech signals and target clean speech signals are fed into the NN for training. With sufficient training, the NN will able to remove background noise and thus improve the quality of speech during the testing stage. In this project, MATLAB software was used to implement the NN. The objective of this software has allowed the user to enhance noisy speech signals and measure both Signal-to-Noise (SNR) ratio and Segmental SNR (SEGSNR). Median filter (mfilter) was introduced to the NN after network training to increase the SNR and SEGSNR values by filtering off "shot" or impulse noise that existed in the speech. Experiments and anallysis were carried out to discover the ideal NN which produces the best speech enhancement results. Last but not least, problems faced and recommended further works was discussed to improve the current Time-Domain Neural Network.