Deep Learning-Based Speech Enhancement Algorithm Using Charlier Transform

Machine learning, a part of artificial intelligence, is recently used in speech enhancement algorithms (SE). The primary focus of SE is finding the original speech signal from the distorted one. Specifically, deep learning is used in SE because it handles nonlinear mapping problems for complicated features. In this paper, Charlier polynomials-based discrete transform, simply discrete Charlier transform (DCHT), has been used to get the spectra of the noisy signal using a fully connected neural network. Deep learning effectively acquires the context information of speech signal and gets enhanced speech with good quality and intelligibility properties. The proposed algorithm is tested experimentally through self-comparison to obtain the best speech enhancement models corresponding to the DCHT parameter. The experiment is performed with different values of the DCHT parameter. In addition, the well-known TIMIT database is used for evaluation purposes. Different speech measures are used in the experiment. The realized results show the ability of the trained model based on DCHT to enhance the speech signal and provide good results on specific conditions.