Fast speech enhancement using a novel noise constrained least square estimation

This paper proposes a fast speech enhancement algorithm for the removal of noise from single-channel speech signal, based on a novel noise constrained least-squares (NCLS) method. Parameters of speech signal modeled as autoregressive process are well estimated by the NCLS method and thus the speech signal can be recovered from Kalman filtering. Simulation results show that the proposed NCLS estimation-based algorithm has a much faster speed than the generalized least absolute deviation estimation-based algorithm and possesses good speech enhancement performance than the Kalman filtering algorithms based on the conventional second-order estimation and the high-order estimation.