Residual Echo Suppression Based on Estimation of Acoustic Echo Path Using Echo Replica

Abstract

This paper presents and compares algorithms for acoustic echo and noise suppression based on a statistical model for hands-free telecommunications. The proposed RES postfilter modifying the estimation of frequency response of acoustic echo path and applying soft decision scheme to the gain filter. Specifically, we estimate the residual echo by the improved least square estimator with echo replica using the spectral correlation between the residual echo and the echo replica. In addition, the MMSE-LSA filter is applied to the gain filter of the proposed algorithm combined with the soft decision considering the probability of near-end speech presence. In the evaluation, we could demonstrate that the proposed algorithm performed significantly better for removing the residual echo and noise without speech distortion compared with the conventional AES and RES algorithms.