Time-frequency masking based supervised speech enhancement framework using fuzzy deep belief network

In recent years, deep learning based supervised speech enhancement methods have gained a considerable amount of research attention over the statistical signal processing based methods. In this study, we have considered the time-frequency masking based deep learning framework for speech enhancement and investigated how the performance of these methods can be improved further. We have mainly established that significant performance improvement can be achieved if the deep neural network (DNN) is pre-trained by using Fuzzy Restricted Boltzmann Machines (FRBM) rather than using regular Restricted Boltzmann Machines (RBM). This is mainly because of the fact that the performance of FRBM is more robust and effective when the training data is noisy. In order to train an FRBM, we have adopted a defuzzification method based on the crisp probabilistic mean value of fuzzy numbers. The detailed theory of training strategy of an FRBM with different fuzzy membership functions such as Symmetric Triangular Fuzzy Numbers (STFN) and Asymmetric Triangular Fuzzy Numbers (ATFN) is presented. Furthermore, we have evaluated the performance of the proposed training strategies on different DNN based Speech Enhancement Systems (SES) which are developed based on different training targets such as Complex Ideal Ratio Mask (cIRM), Ideal Ratio Mask (IRM) and Phase-Sensitive Mask (PSM). Experimental results on various noise scenarios have shown that the DNN-based speech enhancement system trained by the proposed approach ensures a consistent improvement in various objective measure scores of perceived speech quality and intelligibility while compared to the conventional DNN-based speech enhancement methods which use regular RBM for unsupervised pre-training.