A study on noise estimation based on robust equation error IIR ADF for speech

In this paper, we propose the noise estimation method based on system identification with an equation error infinite impulse response (IIR) adaptive digital filter (ADF). We have investigated the speech enhancement based on a linear prediction error filter (LPEF) and a system identification model. However, the conventional system estimates the background noise by a finite impulse response (FIR) filter, therefore, a kind of background noise, which can be estimated, is limited. In addition, the estimation accuracy of the background noise is deteriorated due to disturbance in a speech section. Thus, the robust IIR ADF is introduced to noise estimation. In case that an output error IIR ADF is used, the system has the problem of stability. On the other hand, though the equation error IIR ADF guarantees system stability, the tap coefficients converge with a bias due to disturbance. In order for an adaptive algorithm to converge without a bias, the equation error IIR ADF, which uses the correlation as an input signal, is adopted. The IIR ADF takes advantage of independence between a desired signal and disturbance. Since the disturbance does not have the correlation from the desired signal, the tap coefficients converge without the bias.