Speech enhancement based on bias free noise reconstruction method

In this paper, a speech enhancement method to reduce background noise in noisy speech is proposed. We have investigated the noise reconstruction system (NRS) based on linear prediction and system identification. First, a liner prediction error filter (LPEF) estimates the white noise which is a noise source. Next, a noise reconstruction filter (NRF) estimates the background noise from the estimated white noise. However, the conventional system uses finite impulse response (FIR) filters. The estimated white noise therefore contains the residual speech. As a result, the estimation accuracy of background noise is degraded at the NRF. In order to solve the problems, we introduce a lattice filter and an equation error adaptive digital filter (ADF) as the LPEF and the NRF respectively. Since a lattice filter approximates a vocal-tract filter for the speech production process, the residual speech is reduced. An equation error ADF with bias free is used for improving the quality of enhanced speech. However, the bias free equation error ADF degrades the estimation accuracy of background noise, thus the sub-filter is herein introduced to improve its estimation accuracy.