Speech enhancement of spectral magnitude bin trajectories using Gaussian mixture-model based minimum mean-square error estimators

Abstract

Gaussian mixture-model based minimum mean-square error estimators have been applied to speech enhancement in the temporal, transform (e.g., discrete cosine transform), and subspace domains. In this paper, we propose a method for applying a GMM-based MMSE estimator to spectral magnitude-bin trajectories. In addition, methods for incorporating speech presence uncertainty into the proposed system to improve performance are discussed. The proposed system outperforms previously published GMM-based estimators, and the well-known Ephraim and Malah estimator for 8 kHz telephone-quality speech.