Noisy speech enhancement based on improved minimum statistics incorporating acoustic environment-awareness

In this paper, we propose a novel speech enhancement technique based on an improved minimum statistics (MS) approach incorporating acoustic environmental noise awareness. A relevant noise estimation approach, known as MS, tracks the minimal values if a smoothed power estimate of the noisy signal is within a finite search window. From an investigation of previous MS-based methods, it is discovered that a fixed size of the minimum search window is assumed regardless of the environmental conditions. To overcome this limitation, we initially determine the optimal window sizes in terms of the perceived speech quality according to a variety of noise types. We then assign a different search window size according to the determined noise type, for which we use a real-time noise classification algorithm based on the Gaussian mixture model (GMM). The performance of the proposed approach is evaluated by a quantitative comparison method and by objective tests under various noise environments. It was found to yield better results compared to the previous MS method.