Regression-based speech enhancement by convolutional neural network

In this study, a regression-based convolutional neural network (CNN) model is proposed for speech enhancement. The main purpose is to remove the noise on the conversations. A babble noise is added to the speech samples of different persons and samples with different signal to noise ratio (SNR) are obtained. The logarithmic power spectrum (LPS) coefficients of noisy and clean speech signal samples are calculated. Then a regression model is established between the convolutional neural network and the logarithmic power spectrum coefficients of noisy and clean speech. The results are evaluated by perceptual evaluation of speech quality (PESQ) and short time objective intelligibility (STOI). The results are presented in tabular form.