Speech Enhancement with Deep Neural Networks Using MoG Based Labels

In this paper we present a mixture of Gaussians-deep neural network (MoG-DNN) algorithm for single-microphone speech enhancement. We combine between the generative mixture of Gaussians (MoG) model and the discriminative deep neural network (DNN). The proposed algorithm consists of two phases, the training phase and the test phase. In the training phase, the clean speech power spectral density (PSD) is modeled as a MoG representing an unsupervised assortment of the speech signal. Following, the database is labeled to fit the given MoG. DNN is then trained to classify noisy time-frame features to one of the Gaussians from the already inferred MoG. Given the classification results, a speech presence probability (SPP) is obtained in the test phase. Using the SPP, soft spectral subtraction is then applied, while, simultaneously updating the noise statistics. The generative unsupervised MoG can be applied to any unknown database, in addition to preserving the speech spectral structure. Furthermore, the discriminative DNN maintains the continuity of the speech. Experimental study shows that the proposed algorithm produces higher objective measurements scores compared to other speech enhancement algorithms.