A combined statistical and machine learning approach for single channel speech enhancement

Abstract

In this thesis, we study the single-channel speech enhancement problem, the goal of which is to recover a desired speech from a monaural noisy recording. Speech enhancement is a focal issue to study due to is widespread usage in speech-related applications, such as hearing aids, mobile communications, and speech recognition systems. Three speech enhancement algorithms are proposed. In the rst algorithm, the Wiener Non-negative Matrix Factorization (WNMF), we combine the traditional Wiener ltering and the NMF into a single optimization problem. The objective is to minimize the mean square error, similar to Wiener ltering, and the constraints ensure the enhanced speeches are sparsely representable by the speech model learned by NMF. WNMF is novel because it utilizes NMF to capture the speech-specific structure while simultaneously leveraging it, thus improving the Wiener filtering. For the second algorithm, we propose a Sparse Gaussian Mixture Model (SGMM) that extends the traditional NMF and the Gaussian model. SGMM better captures the complex structure of speech than the traditional NMF. To control for overrepresentation of SGMM, we impose sparsity in order to ensure that only a few Gaussian models are simultaneously active. Computationally, it is achieved by using a l0-norm in the constraint of the maximum-likelihood (ML) estimation. The contribution of SGMM is in solving the constrained ML estimation, which has a closed form update even with the non-convex and non-smooth l0-norm constraint. The final algorithm proposed is the Sparse NMF + Deep Neural Network (SNMF-DNN), in which we treat speech enhancement as a supervised regression problem - the goal being to estimate the optimal enhancement gain. SNMF, originally designed for source separation, is used to extract features from the noisy recording. DNN is subsequently trained to estimate the optimal enhancement gain. Although our system is simple and does not require any sophisticated handcrafted features, we are able to demonstrate a substantial improvement in both intelligibility and enhanced speech quality.