Single-Channel Noise Spectral Estimation Based on Compensated Speech Presence Probability

Abstract

Recently, various noise spectrum estimation techniques have been proposed. The problem of the conventional techniques has the underestimation of the noise spectrum due to the excessive guess of the speech presence probability. Therefore, we propose a highly accurate noise spectrum estimation method to solve this problem by applying a sigmoid or softplus functions to the speech presence probability. From the evaluation experiment, we confirm that the proposed methods can achieve higher noise estimation accuracy than conventional techniques.